This page documents the current and known upcoming maintenance windows.
+We will perform OS upgrades on both login nodes during this week. To keep disruption of service as minimal as possible +upgrades will be staggered:
+hpc-login-2 will be offline on Monday 9thhpc-login-1 will be offline on Thurday 12thPlease plan accordingly.
Kernel and Slurm upgrades. Keep an eye on our forum for updates.
~/.bashrc Guide
- December 5, 2024
+ December 9, 2024
diff --git a/best-practice/env-modules/index.html b/best-practice/env-modules/index.html
index 9b2532ee..3d0fefb6 100644
--- a/best-practice/env-modules/index.html
+++ b/best-practice/env-modules/index.html
@@ -3374,7 +3374,7 @@ critical
- December 5, 2024
+ December 9, 2024
diff --git a/overview/monitoring/index.html b/overview/monitoring/index.html
index dedb91a4..c211bd79 100644
--- a/overview/monitoring/index.html
+++ b/overview/monitoring/index.html
@@ -3122,7 +3122,7 @@ Welcome to the user documentation of the BIH high-performance computing (HPC) cluster, also called HPC 4 Research. The BIH HPC cluster is managed by CUBI (Core Unit Bioinformatics). This documentation is maintained by BIH CUBI and the user community. It is a living document that you can update and add to. See How-To: Contribute to this Document for details.
The global table of contents is on the left, the one of the current page is on the right.
Additional resources
Read the following set of pages (in order) to learn how to get access and connect to the cluster.
Acknowledging BIH HPC Usage
Acknowledge usage of the cluster in your manuscript as \"Computation has been performed on the HPC for Research/Clinic cluster of the Berlin Institute of Health\". Please add your publications using the cluster to this list.
"},{"location":"#news-maintenance-announcements","title":"News & Maintenance Announcements","text":"hpc-mem-5 with 4 TB of RAM./fast on all non-transfer nodes.See Maintenance for a detailed list of current, planned, and previous maintenance and update work.
"},{"location":"#technical-details","title":"Technical Details","text":"If you are interested in how this HPC cluster is set up on a technical level, we got you covered. There is an entire section on this.
"},{"location":"#documentation-structure","title":"Documentation Structure","text":"The documentation is structured as follows:
Access to the BIH HPC cluster is conceptually based on user groups (also known as labs or units) and projects. Users have a relatively limited storage quota within their private home folder and store big data primarily within their group's work space or in project folders. Projects are collaborative efforts involving multiple PIs/groups and are allocated separate storage space on the cluster.
Independent group leaders at BIH/Charit\u00e9/MDC can request a group on the cluster and name group members. The work group leader (the group PI) bears the responsibility for the group's members and ensures that cluster policies and etiquette are followed. In brief: Fair usage rules apply and the cluster ist not to be abused for unethical or illegal purposes. Major and/or continued violations may lead to exclusion of the entire group.
The group leader may also name one delegate (typically an IT-savvy Post-Doc) who is thereby allowed to take decision about cluster usage and work group management on behalf of the group leader. The above mentioned responsibilities stay with the group leader.
Note
All cluster users are member of exactly one primary work group. This affiliation is usually defined by real life organisational structures within Charit\u00e9/BIH/MDC. Leaders of independent research groups (PIs) can apply for a new cluster work group as follows:
Important
Changes to an existing group (adding new users, changes in resources, etc.) can only be requested by group leaders and delegates.
"},{"location":"admin/getting-access/#form-new-group","title":"Form: New Group","text":"Example values are given in curly braces.
# Group \"ag-{doe}\"\nGroup leader/PI: {John Doe}\nDelegate [optional]: {Max Mustermann}\nPurpose of cluster usage [short]: {RNA-seq analysis in colorectal cancer}\n\nRequired resources:\n- Tier 1 storage: {1 TB}\n- Tier 1 scratch: {10 TB}\n- Tier 2 storage: {10 TB}\n\n# Users\n## User 1\n- first name: {John}\n- last name: {Doe}\n- affiliation: {Charit\u00e9, Department of Oncology}\n- institute email: {john.doe@charite.de}\n- user has account with\n - [ ] BIH\n - [x] Charite\n - [ ] MDC\n- BIH/Charit\u00e9/MDC user name: {doej}\n\n## User 2\n[etc.]\nExample values are given in curly braces.
# New user of AG {Doe}\n- first name: {Mia}\n- last name: {Smith}\n- affiliation: {Charit\u00e9, Department of Oncology}\n- institute email: {mia.smith@charite.de}\n- user has account with\n - [ ] BIH\n - [x] Charite\n - [ ] MDC\n- BIH/Charit\u00e9/MDC user name: {smithm}\nNotes
Projects are secondary user groups to enable:
Project creation can be initiated by group leaders and group delegates as follows:
Important
Changes to an existing project (adding new users, changes in resources, etc.) can only be requested by project owners and delegates. Please send us cluster user names for adding new project members.
"},{"location":"admin/getting-access/#form","title":"Form","text":"Example values are given in curly braces.
# Project \"{doe-dbgap-rna}\"\nProject owner: {John Doe}, {doej_c}\nDelegate [optional]: {Max Mustermann}, {musterm_c}\nPurpose of cluster usage [short]: {RNA-seq data from dbGAP}\n\nRequired resources:\n- Tier 1 work: {0 TB}\n- Tier 1 scratch: {0 TB}\n- Tier 2 storage: {1 TB}\n\nAdditional members (cluster user names):\n- {sorgls_c}\n- ...\nNotes
This page documents the current and known upcoming maintenance windows.
"},{"location":"admin/maintenance/#login-compute-maintenance-october-22nd-2024","title":"Login & Compute maintenance, October 22nd, 2024","text":"Kernel and Slurm upgrades. Keep an eye on our forum for updates.
"},{"location":"admin/maintenance/#login-compute-and-storage-maintenance-december-13-14-2022","title":"Login, Compute and Storage Maintenance, December 13-14, 2022","text":"All informationand updates regarding maintenance will be circulated on our forum https://hpc-talk.cubi.bihealth.org/c/announcements/5.
"},{"location":"admin/maintenance/#login-compute-and-storage-maintenance-march-22-23-2022","title":"Login, Compute and Storage Maintenance, March 22-23, 2022","text":"All COMPUTE nodes and STORAGE resources won't be reachable!
All nodes will be running in RESERVATION mode. This means you are still able to schedule new jobs on these nodes if their potential/allowed runtime does not extend into the maintenance window (Tuesday and Wednesday, March 22 and 23, all-day). For example, if you submit a job that can run up to 7 days after March 15 then the job will remain in \"pending/PD\" state giving the explanation of \"all nodes being reserved or unavailable\".
Issues of today's maintenance:
/tmp on login nodescephfs-2 switches (Tier 2 storage, not relevant for most users)IMPORTANT
Progress Thread on hpc-talk
"},{"location":"admin/maintenance/#drmaa-deprecation-march-2-2022","title":"DRMAA Deprecation, March 2, 2022","text":"scontrol show job JOBID and sacct -j JOBID.snakemake --profile=cubi-v1 instead of snakemake --drmaa \"...\".rule myrule:\n # ...\n threads: 8\n resources:\n time=\"12:00:00\",\n memory=\"8G\",\n # ...\nSchedulerParameters+=bf_max_job_user=50: backfill scheduler only considers 50 jobs of each user. This mitigates an issue with some users having too many jobs and thus other users' jobs don't get ahead in the queueEnforcePartLimits=ALL: jobs that don't fit into their partition are rejectedDependencyParameters=kill_invalid_depend: jobs that have dependencies set that cannot be fulfilled will be killedlocaltmp Resource, January 31, 2022","text":"localtmp resource for local storage above 100MB./tmp using Linux namespaces/cgroups. This greatly improves the reliability of cleaning up after jobs. (Technically, this is implemented using the Slurm job_container/tmpfs) plugin.Gres) \"localtmp\". In the future this will become a requirement. Also see Slurm: Temporary Files.hpc-login-1.cubi.bihealth.orghpc-login-2.cubi.bihealth.orghpc-portal.cubi.bihealth.orghpc-transfer-1.cubi.bihealth.orghpc-transfer-2.cubi.bihealth.orghpc-gpu-{5..7}. 28.08.5.The GPFS storage system has been upgraded to the latest version to make compatible with Enterprise Linux version 8.
"},{"location":"admin/maintenance/#slurm-upgrade-to-21080-september-8-2021","title":"Slurm upgrade to21.08.0, September 8, 2021","text":"Slurm has been upgraded to version 21.08.0.
All servers/nodes won't be reachable!
All nodes will be running in reservation mode. This means you are still able to schedule new jobs on these nodes if their potential/allowed runtime does not extend into the maintenance window (Tuesday and Wednesday, September 7 and 8, all-day). For example, if you submit a job that can run up to 7 days after August 30 then the job will remain in \"pending/PD\" state giving the explanation of \"all nodes being reserved or unavailable\".
If you already have a job running on any nodes that goes beyond September 7, 12:00 am (00:00 Uhr), this job will die.
"},{"location":"admin/maintenance/#renaming-of-gpu-high-memory-machines-scheduler-changes-september-7-2021","title":"Renaming of GPU & High Memory Machines & Scheduler Changes, September 7, 2021","text":"The GPU machines med030[1-4] have been renamed to hpc-gpu-[1-4]. The high memory machines med040[1-4] have been renamed to hpc-mem-[1-4]. It will probably take us some time to update all places in the documentation.
Further, the long partition has been changed to allow jobs with a maximum running time of 14 days.
staging partition, August 31, 2021","text":"We have installed 36 new nodes (in BETA mode) in the cluster called hpc-node-[1-36]. They have 48 cores (thus 96 hardware threads) each and have 360GiB of main memory available (for the hardware nerds, it's Intel(R) Xeon(R) Gold 6240R CPUs at 2.40GHz, featuring the cascadelake architecture).
Right now, they are only available in the staging partition. After some testing we will move them to the other partitions. We'd like to ask you to test them as well and report any issues to hpc-helpdesk@bih-charite.de. The nodes have been setup identically to the existing med0xxx nodes. We do not expect big changes but the nodes might not be as stable as other oness.
Here is how you can reach them.
hpc-login-1 # srun --immediate=5 --pty --time=24:00:00 --partition=staging bash -i\n[...]\nhpc-cpu-1 #\nNote that I'm specifying a maximal running time of 24h so the scheduler will end the job after 24 hours which is before the upcoming maintenance reservation begins. By default, the scheduler allocates 28 days to the job which means that the job cannot end before the reservation and will be scheduled to start after it. See Reservations / Maintenances for more information about maintenance reservations.
"},{"location":"admin/maintenance/#reservation-maintenance-display-on-login-august-30-2021","title":"Reservation / Maintenance Display on Login, August 30, 2021","text":"User will now be notified on login about maintenance, for example:
NOTE: scheduled maintenance(s)\n\n 1: 2021-09-07 00:00:00 to 2021-09-09 00:00:00 ALL nodes\n\nSlurm jobs will only start if they do not overlap with scheduled reservations.\nMore information:\n\n - https://bihealth.github.io/bih-cluster/slurm/reservations/\n - https://bihealth.github.io/bih-cluster/admin/maintenance/\nThe srun command will now behave as if --immediate=60 has been specified by default. It explains how to override this behaviour and possible reasons for job scheduling to fail within 60 seconds (reservations and full cluster).
We upgrade from 20.11.2 to 20.11.8 which contains some fixes for bugs that our users actually stumbled over. The change should be non-intrusive as it's only a patch-level update.
Following servers won't be reachable:
These nodes are running in reservation mode now. This means you are still able to schedule new jobs on these nodes if their potential/allowed runtime does not extend into the maintenance window (Tuesday, August 3, all-day). For example, if you submit a job that can run up to 7 days after July 26 then the job will remain in \"pending/PD\" state giving the explanation of \"all nodes being reserved or unavailable\". If you have a job running on any of the before mentioned nodes that goes beyond August 3, 12:00 am (00:00 Uhr), this job will die. We do not expect the remaining nodes to be affected. However, there remains a minor risk of unexpected downtime of other nodes.
"},{"location":"admin/maintenance/#server-reorganization-july-13-2021","title":"Server reorganization, July 13, 2021","text":"Affected servers are:
If you have a job running on any of the before mentioned nodes that goes beyond June 22, 6am, this job will die. We put a so-called Slurm reservation for the maintenance period. Any job that is scheduled before the maintenance and whose end time (start time + max running time) is not before the start of the maintenance will not be scheduled with the message ReqNodeNotAvail, Reserved for maintenance.
Affected servers are:
HPC 4 Research
Note
This task is currently being planned. No schedule has been fixed yet.
Note
This task is currently being planned. No schedule has been fixed yet.
/fast that currently points to /data/gpfs-1 on HPC 4 Research./data instead of /fast everywhere, e.g., /data/users/$NAME etc.Time: 6am-12am
/fast file system will be re-mounted to /data/gpfs-1./fast becomes a symbolic link to /data on all of the cluster.hpc-login-1.cubi.bihealth.org and login-2... instead of hpc-login-{1,2}.hpc-transfer-{1,2} which will be replaced by transfer-1.research.hpc.bihealth.org and transfer-2....med010[1-3] and med012[5-6].On June 3, we need to perform a network maintenance at 8 am.
If everything goes well, there might be a short delay in network packages and connections will survive. In this case, the maintenance will end 8:30 am.
Otherwise, the maintenance will finish by noon.
"},{"location":"admin/maintenance/#cluster-maintenance-with-downtime-june-16","title":"Cluster Maintenance with Downtime: June 16","text":"We need to schedule a full cluster downtime on June 16.
"},{"location":"admin/maintenance/#slurm-migration","title":"Slurm Migration","text":"We will switch to the Slurm workload scheduler (from the legacy SGE). The main reason is that Slurm allows for better scheduling of GPUs (and has loads of improvements over SGE), but the syntax is a bit different. Currently, our documentation is in an transient state. We are currently extending our Slurm-specific documentation.
SSH Key Management has switched to using Charite and MDC ActiveDirectory servers. You need to upload all keys by the end of April 2020.
Schedule
Feb 4, 2020: Keys are now also taken from central MDC/Charite servers. You do not need to contact us any more to update your keys (we cannot accelerate the process at MDC).May 1, 2020: Keys are now only taken from central MDC/Charite servers. You must upload your keys to central servers by then.Affected systems:
hpc-transfer-1hpc-transfer-2hpc-login-2The compute nodes are non-critical as we are taking them out of the queues now.
"},{"location":"admin/maintenance/#centos-76-upgrade-january-29-february-5","title":"CentOS 7.6 Upgrade, January 29, February 5","text":"Starting monday 03.09.2018 we will be performing rolling update of the cluster from CentOS 7.4 to CentOS 7.5. Since update will be performed in small bunches of nodes, the only impact you should notice is smaller number of nodes available for computation.
Also, for around two weeks, you can expect that your jobs can hit both CentOS 7.4 & CentOS 7.5 nodes. This should not impact you in any way, but if you encounter any unexpected behavior of the cluster during this time, please let us know.
At some point we will have to update the transfer, and login nodes. We will do this also in parts, so the you can switch to the other machine.
Key dates are:
18.09.2018 - hpc-login-1 & hpc-transfer-1 will not be available, and you should switch to hpc-login-2 & hpc-transfer-2 respectively.
25.09.2018 - hpc-login-2 & hpc-transfer-2 will not be available, and you should switch to hpc-login-1 & hpc-transfer-1 respectively.
Please also be informed that non-invasive maintenance this weekend which we announced has been canceled, so cluster will operate normally.
In case of any concerns, issues, do not hesitate to contact us via hpc-admin@bih-charite.de, or hpc-helpdesk@bih-charite.de.
"},{"location":"admin/maintenance/#june-18-2018-0600-1500","title":"June 18, 2018, 0600-1500","text":"Due to tasks we need to perform on BIH cluster, we have planned maintenance:
During maintenance we will perform several actions:
During maintenance whole cluster will not be usable, this includes:
Maintenance window is quite long, since we are dependent on external vendor. However, we will recover services as soon as possible.
We will keep you posted during maintenance with services status.
"},{"location":"admin/maintenance/#march-16-18-2018-mdc-it","title":"March 16-18, 2018 (MDC IT)","text":"MDC IT has a network maintenance from Friday, March 16 18:00 hours until Sunday March 18 18:00 hours.
This will affect connections to the cluster but no connections within the cluster.
"},{"location":"admin/maintenance/#january-17-2018-complete","title":"January 17, 2018 (Complete)","text":"STATUS: complete
The first aim of this window is to upgrade the cluster to CentOS 7.4 to patch against the Meltdown/Spectre vulnerabilities. For this, the login and transfer nodes have to be rebooted.
The second aim of this window is to reboot the file server to mitigate some NFS errors. For this, the SGE master has to be stopped for some time.
"},{"location":"admin/maintenance/#planprogress","title":"Plan/Progress","text":"(since January 2010)
This page describes strictly enforced policies valid on the BIH HPC clusters.
The aim of the HPC systems is to support the users in their scientific work and relies on their cooperation. First and foremost, the administration team enforces state of the art IT security and reliability practices through their organizational and operational processes and actions. We kindly ask user to follow the Cluster Etiquette describe below to allow for fair use and flexible access to the shared resources. Beyond this, policies are introduced or enforced only when required to ensure non-restrictive access to the resources themselves. Major or recurrent breaches of policies may lead to exclusion from service.
We will update this list of policies over time. Larger changes will be announced through the mailing list.
"},{"location":"admin/policies/#cluster-etiquette","title":"Cluster Etiquette","text":"getent passswd $USER to find out the user's office contact details).conda, archive management tools such as tar, (un)zip, or gzip. You should probably only run screen/tmux and maybe a text editor there.hpc-transfer-1 and hpc-transfer-2.In the case of violations marked with a shield () administration reserves the right to remove write and possibly read permission to the given locations. Policies marked with a robot () are automatically enforced.
home, work, and scratch volume). You can request an increase by an email to hpc-helpdesk@bih-charite.de for groups and projects.home 10k files, 1GB spacework 2M files, 1TB spacescratch 20M files, 200TB spacehpc-users and mode is u=rwx,go=; POSIX ACLs are prohibited. This policy is automatically enforced every 5 minutes.u=rwx,g=rwxs,o=; POSIX ACLs are prohibited. This policy is automatically enforced every 5 minutes.scratch/BIH_TRASH after 14 days (by mtime) over night. Trash directories will be removed after 14 further days.touch on files in scratch and subsequently bumping the mtime./tmp). In the case that users need to delete files that they can access but not update/delete, administration will either give write permissions to the Unix group of the work group or project or change the owner to the owner/delegate of this group. This can occur in a group/project directory of a user who has left the organization. In the case that a user leaves the organization, the owner/delegate of the hosting group can request getting access to the user's files with the express agreement of this user./tmp in Slurm-controlled jobs. This will enforce that Slurm can clean up after you.Network connections are a topic important in security. In the case of violations marked with a shield () administration reserves the right to terminate connections without notice and perform other actions.
screen and tmux are only allowed to run on the head nodes. They will be terminated automatically on the compute nodes.srun).~/.ssh/authorized_keys file but their usage is discouraged.~/.bashrc Guide","text":"You can find the current default content of newly created user homes in /etc/skel.bih:
hpc-login-1:~$ head /etc/skel.bih/.bash*\n==> /etc/skel.bih/.bash_logout <==\n# ~/.bash_logout\n\n==> /etc/skel.bih/.bash_profile <==\n# .bash_profile\n\n# Get the aliases and functions\nif [ -f ~/.bashrc ]; then\n . ~/.bashrc\nfi\n\n# User specific environment and startup programs\n\nPATH=$PATH:$HOME/.local/bin:$HOME/bin\n\n==> /etc/skel.bih/.bashrc <==\n# .bashrc\n\n# Source global definitions\nif [ -f /etc/bashrc ]; then\n . /etc/bashrc\nfi\n\n# Uncomment the following line if you don't like systemctl's auto-paging feature:\n# export SYSTEMD_PAGER=\nThis document contains a few tips for helping you using environment modules more effectively. As the general online documentation is lacking a bit, we also give the most popular commands here.
"},{"location":"best-practice/env-modules/#how-does-it-work","title":"How does it Work?","text":"Environment modules are descriptions of software packages. The module command is provided which allows the manipulation of environment variables such as PATH, MANPATH, etc., such that programs are available without passing the full path. Environment modules also allow specifying dependencies between packages and conflicting packages (e.g., when the same binary is available in two packages). Further, environment variables allow the parallel installation of different software versions in parallel and then using software \"a la carte\" in your projects.
List currently loaded modules:
$ module list\nShow all available modules
$ module avail\nLoad one module, make sure to use a specific version to avoid ambiguities.
$ module load Jannovar/0.16-Java-1.7.0_80\nUnload one module
$ module unload Jannovar\nUnload all modules
$ module purge\nGet help for environment modules
$ module help\nGet help for a particular environment module
$ module help Jannovar/0.16-Java-1.7.0_80\nYou can also create your own environment modules. Simply create a directory with module files and then use module use for using the modules from the directory tree.
$ module use path/to/modules\n-bash: module: command not found?","text":"On the login nodes, the module command is not installed. You should not run any computations there, so why would you need environment modules there? ;)
meg-login2$ module\n-bash: module: command not found\nUse srun --pty bash -i to get to one of the compute nodes.
You will certainly finding yourself using a set of programs regularly without it being part of the core cluster installation, e.g., SAMtools, or Python 3. Just putting the appropriate module load lines in your ~/.bashrc will generate warnings when logging into the login node. It is thus recommended to use the following snippet for loading modules automatically on logging into a compute node:
case \"${HOSTNAME}\" in\n login-*)\n ;;\n *)\n # load Python3 environment module\n module load Python/3.4.3-foss-2015a\n\n # Define path for temporary directories, don't forget to cleanup!\n # Also, this will only work after /fast is available.\n export TMPDIR=/data/cephfs-1/home/users/$USER/scratch/tmp\n ;;\nesac\nUnder Construction
This guide was written for the old GPFS file system and is in the process of being updated.
"},{"location":"best-practice/project-structure/#general-aims","title":"General Aims","text":"Mostly, you can separate the files in your projects/pipelines into one of the following categories:
Ideally, scripts and documentation are independent of a given project and can be separated from the rest. Configuration is project-dependent and small and mostly does not contain any sensitive information (such as genotypes that allows for reidentification of donors). In most cases, data might be large and is either also stored elsewhere or together with scripts and configuration can be regenerated easily.
There is no backup of work and scratch
The cluster GPFS file system /fast is not appropriate for keeping around single \"master\" copies of data. You should have a backup and archival strategy for your valuable \"master\" copy data.
In addition, you might need project-specific \"wrapper\" scripts that just call your project-independent script with the correct paths for your project. These scripts rather fall into the \"configuration\" category and should then live together with your configuration.
"},{"location":"best-practice/project-structure/#data","title":"Data","text":"Temporary files
You really should keep temporary files in a temporary directory, set the environment variable TMPDIR appropriately and automatically clean them up (see Useful Tips: Temporary Files)
But how can we put this into practice? Below, we give some examples of how to do this. Note that for simplicity's sake we put all scripts and configuration into one directory/repository contrary to the best practices above. This is for educational purposes only and you should strive for reuseable scripts where it makes sense and separate scripts and configuration.
We will limit this to simple Bash scripts for education's purposes. You should be able to easily adapt this to your use cases.
Thus, the aim is to separate the data from the non-data part of the project such that we can put the non-data part of the project into a separate location and under version control. We call the location for non-data part of the project the home location of your project and the location for the data part of the project the work location of your project.
Overall, we have three options:
Creating the work directory and copy the input files into work/input.
$ mkdir -p project/work/input\n$ cp /data/cephfs-1/work/projects/cubit/tutorial/input/* project/work/input\nCreating the home space. We initialize a Git repository, properly configure the .gitignore file and add a README.md file.
$ mkdir -p project/home\n$ cd project/home\n$ cat <<EOF >.gitignore\n*~\n.*.sw?\nEOF\n$ cat <<EOF >README.md\n# Example Project\n\nThis is an example project with config/scripts linked into work location.\nEOF\n$ git init\n$ git add .gitignore README.md\n$ git commit -m 'Initial project#\nWe then create the a simple script for executing the mapping step and a configuration file that gives the path to the index and list of samples to process.
$ mkdir scripts\n$ cat <<\"EOF\" >scripts/run-mapping.sh\n#!/bin/bash\n\n# Unofficial Bash script mode, see:\n# http://redsymbol.net/articles/unofficial-bash-strict-mode/\nset -euo pipefail\n\n# Get directory to bash file, see\n# https://stackoverflow.com/a/4774063/84349\nSCRIPTPATH=\"$( cd \"$(dirname \"$0\")\" ; pwd -P )\"\n\n# Helper function to print help to stderr.\nhelp()\n{\n >&2 echo \"Run Mapping Step\"\n >&2 echo \"\"\n >&2 echo \"run-mapping.sh [-c config.sh] [-h]\"\n}\n\n# Parse command line arguments into bash variables.\nCONFIG=\nwhile getopts \"hs:\" arg; do\n case $arg in\n h)\n help()\n exit\n ;;\n s)\n CONFIG=$OPTARG\n ;;\n esac\ndone\n\n# Print the executed commands.\nset -x\n\n# Load default configuration, then load configuration file if any was given.\nsource $SCRIPTPATH/../config/default-config.sh\nif [[ -z \"$CONFIG\" ]]; then\n source $CONFIG\nfi\n\n# Create output directory.\nmkdir -p output\n\n# Actually perform the mapping. This assumes that you have\n# made the bwa and samtools commands available, e.g., using conda.\nfor sample in $SAMPLES; do\n bwa mem \\\n $BWA_INDEX \\\n input/${sample}_R1.fq.gz \\\n input/${sample}_R2.fq.gz \\\n | samtools sort \\\n -o output/${sample}.bam \\\n /dev/stdin\ndone\n\nEOF\n$ chmod +x scripts/run-mapping.sh\n$ mkdir -p config\n$ cat <<\"EOF\" >config/default-config.sh\nBWA_INDEX=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/hs37d5/hs37d5.fa\nSAMPLES=\nEOF\n$ cat <<\"EOF\" >config/project-config.sh\n$ BWA_INDEX comes from default configuration already\nSAMPLES=test\nEOF\nThis concludes the basic project setup. Now, to the symlinks:
$ cd ../work\n$ ln -s ../home/scripts ../home/config .\nAnd, to the execution...
$ ./scripts/run-mapping -c config/project-config.sh\n[...]\nWe can reuse the project up to the statement \"This concludes the basic project setup\" in the example for option 1.
Then, we can do the following:
$ cd ../work\n$ mkdir -p output\n\n$ cd ../home\n$ cat <<\"EOF\" >>.gitignore\n\n# Ignore all data\ninput/\nwork/\noutput/\nEOF\n$ git add .gitignore\n$ git commit -m 'Ignoring data file in .gitignore'\n$ ln -s ../work ../output .\nAnd we can execute everything in the home directory.
$ ./scripts/run-mapping -c config/project-config.sh\n[...]\nAgain, we can reuse the project up to the statement \"This concludes the basic project setup\" in the example for option 1.
Then, we do the following:
$ cd ../work\n$ cat <<\"EOF\" >do-run-mapping.sh\n#!/bin/bash\n\n../home/scripts/run-mapping.sh \\\n -c ../home/config/project-config.sh\nEOF\n$ chmod +x do-run-mapping.sh\nNote that the the do-run.sh script could also go into the project-specific Git repository and be linked into the work directory.
Finally, we can run our pipeline:
$ cd ../work\n$ ./do-run-mapping.sh\n[...]\nThe program screen allows you to detach your session from your current login session. So in case you get disconnected your screen session will stay alive.
Hint
You have to reconnect to screen on the machine that you started it. We thus recommend starting it only on the login nodes and not on a compute node.
"},{"location":"best-practice/screen-tmux/#start-and-terminat-a-screen-session","title":"Start and terminat a screen session","text":"You start a new screen session by
$ screen\n$ exit\nIf you want to detach your screen session press Ctrl+a d
To list all your screen sessions run
$ screen -ls\n\nThere is a screen on:\n 2441.pts-1.med0236 (Detached)\n1 Socket in /var/run/screen/S-kbentel.\nTo reattach a screen session run
$ screen -r screen_session_id\nIf you do not know the screen_session_id you can get it with screen -ls, e.g. 2441.pts-1.med0236 in the example above. You do not have to type the whole screen_session_id only as much as is necessary to identify it uniquely. In case there is only one screen session detached it is enough to run screen -r
Sometimes it is necessary to kill a detached screen session. This is done with the command
$ screen -X -S screen_session_id quit\nIt is possible to have multiple windows in a screen session. So suppose you are logged into a screen session, these are the relevant shortcuts
new win: Ctrl+a c\nnext/previous win: Ctrl+a n/p\nTo terminate a window just enter
$ exit\nHere is a sensible screen configuration. Save it as ~/.screenrc.
screenrc
"},{"location":"best-practice/screen-tmux/#fix-a-broken-screen-session","title":"Fix a broken screen session","text":"In case your screen session doesn't write to the terminal correctly, i.e. the formatting of the output is broken, you can fix it by typing to the terminal:
$ tput smam\nComputer software, or simply software, is a generic term that refers to a collection of data or computer instructions that tell the computer how to work, in contrast to the physical hardware from which the system is built, that actually performs the work. -- Wikipedia: Software
As you will most probably never have contact with the HPC system hardware, everything you interact with on the HPC is software. All of your scripts, your configuration files, programs installed by you or administration, and all of your data.
This should also answer the question why you should care about software and why you should try to create and use software of a minimal quality.
Software craftsmanship is an approach to software development that emphasizes the coding skills of the software developers themselves. -- Wikipedia: Software Craftmanship
This Wiki page is not mean to give you an introduction of creating good software but rather collect a (growing) list of easy-to-use and high-impact points to improve software quality. Also, it provides pointers to resources elsewhere on the internet.
"},{"location":"best-practice/software-craftmanship/#use-version-control","title":"Use Version Control","text":"Use a version control system for your configuration and your code. Full stop. Modern version control systems are Git and Subversion.
Every user should have their own Git/Subversion checkout. Otherwise you are inviting a large number of problems.
"},{"location":"best-practice/software-craftmanship/#document-your-code","title":"Document Your Code","text":"This includes
Document where you got things from, how to re-download, etc. E.g., put a README file into each of your data top level directories.
"},{"location":"best-practice/software-craftmanship/#use-checksums","title":"Use Checksums","text":"Use MD5 or other checksums for your data. For example, md5sum and hashdeep are useful utilities for computing and checking them:
md5sum How-To (tools such as sha256sum work the same...)hashdeep How-ToUse some system for managing your workflows. These systems support you by
Snakemake is a popular workflow management system widely used in Bioinformatics. A minimal approach is using Makefiles.
"},{"location":"best-practice/software-craftmanship/#understand-bash-and-shell-exit-codes","title":"Understand Bash and Shell Exit Codes","text":"If you don't want to use a workflow management system, e.g., for one-step jobs, you should at least understand Bash job management and exit codes. For example, you can use if/then/fi in Bash together with exit codes to:
if [[ ! -e file.md5 ]]; then\n md5sum file >file.md5 \\\n || rm -f file.md5\nfi\nAlso, learn about the inofficial Bash strict mode.
"},{"location":"best-practice/software-installation-with-conda/","title":"Software Installation with Conda","text":""},{"location":"best-practice/software-installation-with-conda/#conda","title":"Conda","text":"Users do not have the rights to install system packages on the BIH HPC cluster. For the management of bioinformatics software we therefore recommend using the conda package manager. Conda provides software in different \u201cchannels\u201d and one of those channels contains a huge selection of bioinformatics software (bioconda). Generally packages are pre-compiled and conda just downloads the binaries from the conda servers.
You are in charge of managing your own software stack, but conda makes it easy to do so. We will provide you with a description on how to install conda and how to use it. Of course there are many online resources that you can also use. Please find a list at the end of the document.
Warning
Following a change in their terms of service Anaconda Inc. has started to demand payment from research institutions for using both Anaconda, Miniconda, and the defaults channel. As a consequence, usage of this software is prohibited and we're recommending the alternative free \"miniforge\" distribution instead.
"},{"location":"best-practice/software-installation-with-conda/#premise","title":"Premise","text":"When you logged into the cluster, please make sure that you also executed srun to log into a computation node and perform the software installation there.
hpc-login-1:~$ srun --mem=5G --pty bash -i\nhpc-cpu-123:~$ wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-x86_64.sh\nhpc-cpu-123:~$ bash Miniforge3-Linux-x86_64.sh -b -f -p $HOME/work/miniforge\nhpc-cpu-123:~$ eval \"$(/$HOME/work/miniforge/bin/conda shell.bash hook)\"\nhpc-cpu-123:~$ conda init\nhpc-cpu-123:~$ conda config --set auto_activate_base false\nThis will install conda to $HOME/work/miniforge. You can change the path to your liking, but please note that your $HOME folder has limited space. The work subfolder however has a bigger quota. More about this here.
To make bioinformatics software available, we have to add the bioconda channel to the conda configuration:
hpc-cpu-123:~$ conda config --add channels bioconda\nInstalling packages with conda is straight forward:
hpc-cpu-123:~$ conda install <package>\nThis will install a package into the conda base environment. We will explain environments in detail in the next section. To search for a package, e.g. to find the correct name in conda or if it exists at all, issue the command:
hpc-cpu-123:~$ conda search <string>\nTo choose a specific version (conda will install the latest version that is compatible with the current installed Python version), you can provide the version as follows:
hpc-cpu-123:~$ conda install <package>=<version>\nPlease note that new conda installs may ship with a recently update Python version and not all packages might have been adapted. E.g., if you find out that some packages don't work after starting out/upgrading to Python 3.8, simply try to downgrade Python to 3.7 with conda install python=3.7.
Hint
As resolving the dependency tree of an installation candidate can take a lot of time in Conda, especially when you are installing software from an environment.yaml file, an alternative resolver has been presented that you can use to install software into your Conda environment. The time savings are immense and an installation that took more than an hour can be resolved in seconds.
Simply run
hpc-cpu-123:~$ conda install mamba\nWith that, you can install software into your environment using the same syntax as for Conda:
hpc-cpu-123:~$ mamba install <package>\nConda lets you create environments, such that you can test things in a different environment or group your software. Another common use case is to have different environments for the different Python versions. Since conda is Python-based, conflicting packages will mostly struggle with the Python version.
By default, conda will install packages into its root environment. Please note that software that does not depend on Python and is installed in the root environment, is is available in all other environments.
To create a Python 2.7 environment and activate it, issue the following commands:
hpc-cpu-123:~$ conda create -n py27 python=2.7\nhpc-cpu-123:~$ source activate py27\n(py27) hpc-cpu-123:~$\nFrom now on, conda will install packages into the py27 environment when you issue the install command. To switch back to the root environment, simply deactivate the py27 environment:
(py27) hpc-cpu-123:~$ source deactivate py27\nhpc-cpu-123:~$\nBut of course, as Python 2.7 is not supported any more by the Python Software Foundation, you should switch over to Python 3 already!
"},{"location":"best-practice/temp-files/","title":"Temporary Files","text":"Temporary Files and Slurm
See Slurm: Temporary Files for information how Slurm controls access to local temporary storage.
Often, it is necessary to use temporary files, i.e., write something out in the middle of your program, read it in again later, and then discard these files. For example, samtools sort has to write out chunks of sorted read alignments for allowing to sort files larger than main memory.
TMPDIR","text":"Traditionally, in Unix, the environment variables TMPDIR is used for storing the location of the temporary directory. When undefined, usually /tmp is used.
Generally, there are two locations where you could put temporary files:
/data/cephfs-1/home/users/$USER/scratch/tmp -- inside your scratch folder on the CephFS file system; this location is available from all cluster nodes/tmp -- on the local node's temporary folder; this location is only available on the node itself. The slurm scheduler uses Linux namespaces such that every job gets its private /tmp even when run on the same node.scratch/tmp","text":"Use CephFS-based TMPDIR
Generally setup your environment to use /data/cephfs-1/home/users/$USER/scratch/tmp as filling the local disk of a node with forgotten files can cause a lot of problems.
Ideally, you append the following to your ~/.bashrc to use /data/cephfs-1/home/users/$USER/scratch/tmp as the temporary directory. This will also create the directory if it does not exist. Further, it will create one directory per host name which prevents too many entries in the temporary directory.
export TMPDIR=$HOME/scratch/tmp/$(hostname)\nmkdir -p $TMPDIR\nPrepending this to your job scripts is also recommended as it will ensure that the temporary directory exists.
"},{"location":"best-practice/temp-files/#tmpdir-and-the-scheduler","title":"TMPDIR and the scheduler","text":"In the older nodes, the local disk is a relatively slow spinning disk, in the newer nodes, the local disk is a relatively fast SSD. Further, the local disk is independent from the CephFS file system, so I/O volume to it does not affect the network or any other job on other nodes. Please note that by default, Slurm will not change your environment variables. This includes the environment variable TMPDIR.
Slurm will automatically update temporary files in a job's /tmp on the local file system when the job terminates. To automatically clean up temporary directories on the shared file system, use the following tip.
You can use the following code at the top of your job script to set TMPDIR to the location in your home directory and get the directory automatically cleaned when the job is done (regardless of successful or erroneous completion):
# First, point TMPDIR to the scratch in your home as mktemp will use thi\nexport TMPDIR=$HOME/scratch/tmp\n# Second, create another unique temporary directory within this directory\nexport TMPDIR=$(mktemp -d)\n# Finally, setup the cleanup trap\ntrap \"rm -rf $TMPDIR\" EXIT\nWe recommend to use the program MobaXterm on Windows. MobaXterm is a software that allows you to connect to an SSH server, much like PuTTy, but also maintains your SSH key.
Alternative SSH Clients for Windows
For transfering data from/to Windows, we recommand using WinSCP. Install the latest version from here: https://winscp.net/eng/download.php
On the Login screen of WinSCP create a new login by selecting New Site.
Fill in the following parameters:
File protocol: SFTPHost name: hpc-transfer-1.cubi.bihealth.org or hpc-transfer-2.cubi.bihealth.orgUser name: your user nameGo to Advanced > SSH > Authentication > Authentication parameters > Private key file and select your private ssh key file (in .ppk format).
Press Ok then Save.
Press Login to connect. It will ask for your private key passphrase, if you set one up.
If you need to convert your private ssh key file the .ppk format, on the WinSCP login screen go to Tools > PuTTYgen and follow the steps here: https://docs.acquia.com/cloud-platform/manage/ssh/sftp-key/
Click on Session.
Click on SSH.
In Basic SSH settings, enter a hostname (hpc-login-X.cubi.bihealth.org, where X is 1 or 2), check Specify username and enter your username in the textfield. Select the tab Advanced SSH settings, check Use private key and select your private SSH key file (possible choices described with the next to figures).
Select the id_rsa file generated in Linux OR
select the id_rsa.ppk file generated in Windows with MobaXterm.
Afterwards hit the OK button and MobaXterm will connect.
The session will be stored automatically and you can establish new connections later on, or also multiple ones at the same time, if you like.
"},{"location":"connecting/connecting/","title":"Connecting to HPC 4 Research","text":"HPC 4 Research is only available via the Charit\u00e9, MDC, and BIH internal networks. VPN access requires additional measures which are described in Connecting from External Networks.
There are two primary methods for interacting with BIH HPC:
This part of the documentation only described direct console access via SSH. For information regarding the web portal, please read OnDemand Portal. In case you're not familiar with SSH, you should probably start via the web portal or (if you are determined to learn) read through our SSH basics page.
"},{"location":"connecting/connecting/#in-brief","title":"In brief","text":"Follow these steps to connect to BIH HPC via the command line:
Connect to one of the two login nodes.
# Charite Users\n$ ssh user_c@hpc-login-1.cubi.bihealth.org\n$ ssh user_c@hpc-login-2.cubi.bihealth.org\n\n# MDC Users\n$ ssh user_m@hpc-login-1.cubi.bihealth.org\n$ ssh user_m@hpc-login-2.cubi.bihealth.org\nHint
There are two login nodes, hpc-login-1 and hpc-login-2. There are two for redundancy reasons. Please do not perform big file transfers or an sshfs mount via the login nodes. For this purpose, we have hpc-transfer-1 and hpc-transfer-2.
Please also read Advanced SSH for more custom scenarios how to connect to BIH HPC. If you are using a Windows PC to access BIH HPC, please read Connecting via SSH on Windows
Allocate resources on a computation node using Slurm. Do not compute on the login node!
# Start interactive shell on computation node\n$ srun --pty bash -i\nBonus: Configure your SSH client on Linux and Mac or Windows.
tl;dr
SSH-Based Access:
# Interactive login (choose one)\nssh username@hpc-login-1.cubi.bihealth.org\nssh username@hpc-login-2.cubi.bihealth.org\nsrun --pty bash -i\n\n# File Transfer (choose one)\nsftp local/file username@hpc-transfer-1.cubi.bihealth.org:remote/file\nsftp username@hpc-transfer-2.cubi.bihealth.org:remote/file local/file\n\n# Interactive login into the transfer nodes (choose one)\nssh username@hpc-transfer-1.cubi.bihealth.org\nssh username@hpc-transfer-2.cubi.bihealth.org\nYour username for accessing the cluster are composed of your username at your primary organization (Charit\u00e9/MDC) and a suffix:
<Charite username>_c -> doej_c<MDC username>_m -> jdoe_mPlease read Connecting from External Networks
"},{"location":"connecting/connecting/#i-have-problems-connecting","title":"I have problems connecting","text":"Please read Debugging Connection Problems
"},{"location":"connecting/connection-problems/","title":"Debugging Connection Problems","text":"When you encounter problems with the login to the cluster although we indicated that you should have access, depending on the issue, here is a list of how to solve the problem:
"},{"location":"connecting/connection-problems/#im-getting-a-connection-refused","title":"I'm getting a \"connection refused\"","text":"The full error message looks as follows:
ssh: connect to host hpc-login-1.cubi.bihealth.org port 22: Connection refused\nThis means that your computer could not open a network connection to the server.
<DEST>):ifconfig\ntraceroute <DEST>\nipconfig\ntracepath <DEST>\nYou're logging into BIH HPC cluster! (login-1)\n\n ***Your account has not been granted cluster access yet.***\n\n If you think that you should have access, please contact\n hpc-helpdesk@bih-charite.de for assistance.\n\n For applying for cluster access, contact hpc-helpdesk@bih-charite.de.\n\nuser@login-1's password:\nHint
This is the most common error, and the main cause for this is a wrong username. Please take a couple of minutes to read the What is my username?!
If you encounter this message although we told you that you have access and you checked the username as mentioned above, please write to hpc-helpdesk@bih-charite.de, always indicating the message you get and a detailed description of what you did.
"},{"location":"connecting/connection-problems/#im-getting-a-passphrase-prompt","title":"I'm getting a passPHRASE prompt","text":"You're logging into BIH HPC cluster! (login-1)\n\n *** It looks like your account has access. ***\n\n Login is based on **SSH keys only**, if you are getting a password prompt\n then please contact hpc-helpdesk@bih-charite.de for assistance.\n\nEnter passphrase for key '/home/USER/.ssh/id_rsa':\nHere you have to enter the passphrase that was used for encrypting your private key. Read SSH Basics for further information of what is going on here.
"},{"location":"connecting/connection-problems/#i-can-connect-but-i-get-a-password-prompt","title":"I can connect, but I get a passWORD prompt","text":"You're logging into BIH HPC cluster! (login-1)\n\n *** It looks like your account has access. ***\n\n Login is based on **SSH keys only**, if you are getting a password prompt\n then please contact hpc-helpdesk@bih-charite.de for assistance.\n\nuser@login-1's password:\nThis is diffeerent from passPHRASE prompt
Please see I'm getting a passPHRASE prompt for more information.
When you encounter this message during a login attempt, there is an issue with your SSH key. In this case, please connect with increased verbosity to the cluster (ssh -vvv ...) and mail the output and a detailed description to hpc-helpdesk@bih-charite.de.
If you're using macOS and run into problems that your key is not accepted when connecting, please make sure that you're SSH key is properly unlocked in your SSH agent.
"},{"location":"connecting/from-external/","title":"Connecting from External Networks","text":"This page describes how to connect to the BIH HPC from external networks (e.g., another university or from your home). The options differ depending on your home organization and are described in detail below.
Getting Help with VPN and Gateway Nodes
Please note that the VPNs and gateway nodes are maintained by the central IT departments of Charite/MDC. BIH HPC IT cannot assist you in problems with these serves. Authorative information and documentation is provided by the central IT departments as well.
SSH Key Gotchas
You should use separate SSH key pairs for your workstation, laptop, home computer etc. As a reminder, you will have to register the SSH keys with your home IT organization (MDC or Charite). When using gateway nodes, please make sure to use SSH key agents and agent forwarding (ssh flag \"-A\").
Use the following command to perform a proxy jump via the MDC SSH gateway (ssh1 aka jail1) when connecting to a login node. Note that for logging into the jail, the <MDC_USER> is required.
$ ssh -J <MDC_USER>@ssh1.mdc-berlin.de <HPC_USER>@hpc-login-1.cubi.bihealth.org\nNote
Please Note that the cluster login is independent of access to the MDC jail node ssh1.mdc-berlin.de.
You can find the instructions for getting MDC VPN access here in the MDC intranet below the \"VPN\" heading. Please contact helpdesk@mdc-berlin.de for getting VPN access.
Install the VPN client and then start it. Once VPN has been activated you can SSH to the HPC just as from your workstation.
$ ssh user_m@hpc-login-1.cubi.bihealth.org\nAccess to BIH HPC from external networks (including Eduroam) requires a Charit\u00e9 VPN connection with special access permissions.
"},{"location":"connecting/from-external/#general-charite-vpn-access","title":"General Charit\u00e9 VPN Access","text":"You need to apply for general Charit\u00e9 VPN access if you haven't done so already. The form can be found in the Charite Intranet and contains further instructions. Charit\u00e9 IT Helpdesk can help you with any questions.
"},{"location":"connecting/from-external/#zusatzantrag-b","title":"Zusatzantrag B","text":"Special permissions form B is also required for HPC access. You can find Zusatzantrag B in the Charit\u00e9 intranet. Fill it out and send it to the same address as the general VPN access form above.
Once you have been granted VPN access, start the client and connect to VPN. You will then be able to connect from your client in the VPN just as you do from your workstation.
$ ssh jdoe_c@hpc-login-1.cubi.bihealth.org\nAlternative to using Zusatzantrag B, you can also get access to the Charit\u00e9 VDI (Virtual Desktop Infrastructure). Here, you connect to a virtual desktop computer which is in the Charit\u00e9 network. From there, you can connect to the BIH HPC system.
You need to apply for extended VPN access to be able to access the BIH VDI. The form can be found here. It is important to tick Dienst(e), enter HTTPS and as target view.bihealth.org. Please write to helpdesk@charite.de with the request to access the BIH VDI.
When the access has been set up, follow the instructions on client configuration for Windows, after logging in to the BIH VDI.
"},{"location":"connecting/ssh-basics/","title":"SSH Basics","text":""},{"location":"connecting/ssh-basics/#what-is-ssh","title":"What is SSH?","text":"SSH stands for S ecure Sh ell. It is a software that allows to establish a user-connection to a remote UNIX/Linux machine over the network and remote-control it from your local work-station.
Let's say you have an HPC cluster with hundreds of machines somewhere in a remote data-center and you want to connect to those machines to issue commands and run jobs. Then you would use SSH.
"},{"location":"connecting/ssh-basics/#getting-started","title":"Getting Started","text":""},{"location":"connecting/ssh-basics/#installation","title":"Installation","text":"Simply install your distributions openssh-client package. You should be able to find plenty of good tutorials online. On Windows you can consider using MobaXterm (recommended) or Putty.
Let's call your local machine the client and the remote machine you want to connect to the server.
You will usually have some kind of connection information, like a hostname, IP address and perhaps a port number. Additionally, you should also have received your user-account information stating your user-name, your password, etc.
Follow the instructions below to establish a remote terminal-session.
If your are on Linux
Open a terminal and issue the following command while replacing all the <...> fields with the actual data:
# default port\nssh <username>@<hostname-or-ip-address>\n\n# non-default port\nssh <username>@<hostname-or-ip-address> -p <port-number>\nIf you are on windows
Start putty.exe, go into the Session category and fill out the form, then click the Connect button. Putty also allows to save the connection information in different profiles so you don't have to memorize and retype all fields every time you want to connect.
When you connect to a remote machine via SSH, you will be prompted for your password. This will happen every single time you connect and can feel a bit repetitive at times, especially if you feel that your password is hard to memorize. For those who don't want to type in their password every single time they connect, SSH keys are an alternative way of authentication.
Instead if being prompted for a password, SSH will simply use the key to authenticate. As this key file should be device specific, this also increases security of the login process.
You can generate a new key by issuing:
client:~$ ssh-keygen -t ed25519\n\n# 1. Choose file in which to save the key *(leave blank for default)*\n# 2. Choose a passphrase of at least five characters\nAn SSH key consists of two files, one private and one public key. The public key is installed on remote machines and can only be validated with the matching private key, which is stored on client computers. During the login process this is achieved via public-key cryptography.
Traditionally the algorithm used for this was RSA. Recently elliptic curve cryptography has been developed as a more secure and more performant alternative. We recommend the ed25519 type of SSH key.
The security problem with SSH keys is that anyone with access to the private key has full access to all machines that have the public key installed. Loosing the key or getting it compromised in another way imposes a serious security threat. Therefore, it is best to secure the private key with a passphrase. This passphrase is needed to unlock and use the private key.
Once you have your key-pair generated, you can easily change the passphrase of that key by issuing:
client:~$ ssh-keygen -p\nIn order to avoid having to type the passphrase of the key every time we want to use it, the key can be loaded into an SSH-Agent.
For instance, if you have connected to a login-node via Putty and want to unlock your private key in order to be able to access cluster nodes, you cant configure the SSH-Agent.
client:~$ source <(ssh-agent)\n(The above command will load the required environment variables of the SSH-Agent into your shell environment, effectively making the agent available for your consumption.)
Next, you can load your private key:
client:~$ ssh-add\n(You will be prompted for the passphrase of the key)
You can verify that the agent is running and your key is loaded by issuing:
client:~$ ssh-add -l\n# 'l' as in list-all-loaded-keys\n(The command should print at least one key, showing the key-size, the hash of the key-fingerprint and the location of the file in the file-system.)
Since all home-directories are shared across the entire cluster and you created your key-pair inside your home-directory, you public-key (which is also in your home-directory) is automatically installed on all other cluster nodes, immediately. Try connecting to any cluster node. It should not prompt your for a password.
There is nothing you have to do to \"unload\" or \"lock\" the key-file. Simply disconnect.
"},{"location":"connecting/advanced-ssh/overview/","title":"Advanced SSH usage","text":"Here we describe custom scenarios for using SSH to connect to BIH HPC. To keep it consise, this section is divided into separate documents for
Note
The big Linux distributions automatically manage ssh-agent for you and unlock your keys at login time. If this doesn't work for you, read on.
ssh-agent caches your SSH keys so that you do not need to type your passphrase every time it is used. Activate it by making sure ssh-agent runs in the background and add your key:
$ eval \"$(ssh-agent -s)\"\n$ ssh-add\nor if you chose a custom key name, specify the file like so:
$ ssh-add ~/.ssh/mdc_id_rsa\nIf you run into problems that your key is not accepted when connecting from MacOS, please use:
$ ssh-add --apple-use-keychain\nYou can define a personal SSH configuration file to make connecting to the cluster more comfortable by reducing the typing necessary by a lot. Add the following lines to the file ~/.ssh/config file. Replace USER_NAME with your cluster user name. You can also adapt the Host naming as you like.
Host bihcluster\n HostName hpc-login-1.cubi.bihealth.org\n User USER_NAME\n\nHost bihcluster2\n HostName hpc-login-1.cubi.bihealth.org\n User USER_NAME\nNow, you can do type the following (and you don't have to remember the host name of the login node any more).
$ ssh bihcluster\nThis configuration works if you are inside Charit\u00e9, the Charit\u00e9 VPN, or MDC.
"},{"location":"connecting/advanced-ssh/unix/#mdc-users-jail-node","title":"MDC users: Jail node","text":"If you have an MDC user account and want to connect from the outside, you can use the following ~/.ssh/config lines to set up a ProxyJump via the MDC SSH jail.
Host mdcjail\n HostName ssh1.mdc-berlin.de\n User MDC_USER_NAME\nNow you can run
$ ssh -J mdcjail bihcluster1\nIf you are always connecting from outside the internal network, you can also add a permanent ProxyJump to the SSH configuration like so:
Host bihcluster\n HostName hpc-login-1.cubi.bihealth.org\n User USER_NAME\n ProxyJump mdcjail\nIf you need to connect to the cluster from another computer than the one that contains the SSH keys that you submitted for the cluster login, you have two possibilities.
~/.ssh/id_rsa) to the second computer into the same location.Danger
Do not leave the key on any USB stick. Delete it after file transfer. This is a sensible part of data. Make sure that the files are only readable for you.
$ cd ~/.ssh\n$ chmod g-rwx id_rsa*\n$ ssh-add id_rsa\n$ sshfs <USERNAME>@hpc-transfer-1.cubi.bihealth.org:/ <MOUNTPOINT>\nhpc-transfer-1: follows the structure <host>:<directory> starting in the user home.<MOUNTPOINT> must be an empty but existing and readable directory on your local computerMake sure you have both OSXFUSE and SSHFS installed. You can get both from here: https://osxfuse.github.io/ or the most recent version via Homebrew:
$ brew cask install osxfuse; brew install sshfs; brew link --overwrite sshfs\n$ sshfs -o follow_symlinks <USERNAME>@hpc-transfer-1<X>.cubi.bihealth.org:<directory_relative_to_Cluster_root> <MOUNTPOINT> -o volname=<BIH-FOLDER> -o allow_other,noapplexattr,noappledouble\nDo you really need to run a graphical application on the cluster?
Please note that running more complex Java applications, such as IGV may be not very efficient because of the connection speed. In most cases you can run them on your local workstation by mounting them via SSHFS.
Connect to one of the login nodes using X11 forwarding:
$ ssh -X -C -t <USERNAME>@hpc-login-1.bihealth.org\nOnce you get a login prompt, you can use the srun command with the --x11 parameter to open a X11 session to a cluster node:
$ srun --pty --x11 bash\nAnd finally you can start your X11 application, e.g.:
$ xterm\nAfter a while Visual Terminal should start:
"},{"location":"connecting/advanced-ssh/windows/","title":"Windows","text":""},{"location":"connecting/advanced-ssh/windows/#mounting-the-fs-from-within-the-mdccharite-network","title":"Mounting the FS from within the MDC/Charite Network","text":"Danger
Mounting ssh on Windows is currently discouraged since relevant software is outdated (see also hpc-talk). Also, in most cases it is not really necessary to have a constant mount. For normal data transfer please use WinSCP instead.
Once WinSshFS is started, an icon will be added to your taskbar:
Left-clicking that icon will bring up a window. If not, right click the taskbar icon, select Show Manager and click Add in the menu.
Fill out the marked fields:
hpc-transfer-1.cubi.bihealth.orgPrivateKey. Select the id_rsa private key, not the .ppk format that is provided by PuTTY. Enter the password that you used to secure your key with.Then click Save and then Mount.
Open the explorer. A new drive with the name you gave should show up:
Finished!
"},{"location":"connecting/advanced-ssh/windows/#connecting-via-mdc-jail-node","title":"Connecting via MDC Jail Node","text":"This requires an active MDC account!
Additional to the steps above, click on the tab Network settings.
ssh1.mdc-berlin.de and in the field User your MDC username.Do you really need to run a graphical application on the cluster?
Please note that running more complex Java applications, such as IGV may be not very efficient because of the connection speed. In most cases you can run them on your local workstation by mounting them via SSHFS.
Start MobaXterm, it should automatically fetch your saved Putty sessions as you can see on screen below:
Connect to one of the login nodes, by double-click on saved profile, and then use srun --pty --x11 bash command to start X11 session to one of the nodes:
Finally, start X11 application (below example of starting Visual Terminal):
"},{"location":"connecting/generate-key/unix/","title":"Generating an SSH Key with Unix-like Operating Systems","text":"~/.ssh/id_xxx.pub is present.$ ssh-keygen -t ed25519 -C \"your_email@example.com\"\nWhat is a key passphrase?
You should set a passphrase when generating your key pair. It is used for encrypting your private key in case it is stolen or lost. When using the key for login, you will have to enter the passphrase. Many desktop environments offer ways to automatically unlock your key on login.
Read SSH Basics for more information.
The whole session should look something like this:
host:~$ ssh-keygen -t ed25519 -C \"your_email@example.com\"\nGenerating public/private ed25519 key pair.\nEnter file in which to save the key (/home/USER/.ssh/id_ed25519): \nCreated directory '/home/USER/.ssh'.\nEnter passphrase (empty for no passphrase):\nEnter same passphrase again: \nYour identification has been saved in /home/USER/.ssh/id_ed25519.\nYour public key has been saved in /home/USER/.ssh/id_ed25519.pub.\nThe key fingerprint is:\nSHA256:Z6InW1OYt3loU7z14Kmgy87iIuYNr1gJAN1tG71D7Jc your_email@example.com\nThe key's randomart image is:\n+--[ED25519 256]--+\n|.. . . o |\n|. . . + + |\n|. . = . . |\n|. . +oE. |\n|. So= o o |\n| . . . * = + + |\n| + o + B o o .|\n| oo+. .B + + . |\n|.ooooooo*. . |\n+----[SHA256]-----+\nThe file content of ~/.ssh/id_ed25519.pub should look something like this):
ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIFzuiaSVD2j5y6RlFxOfREB/Vbd+47ABlxF7du5160ZH your_email@example.com\nAs a next step you need to submit the SSH key use these links as:
Prerequisite: Installing an SSH Client
Please install an SSH client for Windows first.
"},{"location":"connecting/generate-key/windows/#generate-the-key","title":"Generate the Key","text":"Click on Tools and MobaKeyGen (SSH key generator)
In the section Parameters make sure to set the following properties:
RSA (this is the SSH-2 protocol)4096If all is set, hit the Generate button.
During generation, move the mouse cursor around in the blank area.
When finished, make sure to protect your generated key with a passphrase. Save the private and public key. The default name under Linux for the public key is id_rsa.pub and id_rsa for the private key, but you can name them however you want (the .pub is NOT automatically added). Note that in the whole cluster wiki we will use this file naming convention. Also note that the private key will be stored in Putty format (.ppk, this extension is added automatically).
What is your key's passphrase?
You should set a passphrase when generating your private key. This passphrase is used for encrypting you private key to protect it against the private key file theft/being lost. When using the key for login, you will have to enter it (or the first time you load it into the SSH key agent). Note that when being asked for the passphrase this does not occur on the cluster (and is thus unrelated to it) but on your local computer.
Also see SSH Basics for more information.
The gibberish in the textbox is your public key in the format how it has to be submitted to the MDC and Charite (links for this step below). Thus, copy this text and paste it to the SSH-key-submission-web-service of your institution.
Store the private key additionally in the OpenSSH format. To do so, click Conversions and select Export OpenSSH key. To be consistent, give the file the same name as your .ppk private key file above (just without the .ppk).
To summarize, you should end up with three files:
id_rsa.pub The public key file, it is not required if you copy and submit the SSH public key as described above and in the links below.id_rsa.ppk This file is only needed if you plan to use Putty.id_rsa This is your private key and the one and only most important file to access the cluster. It will be added to the sessions in MobaXterm and WinSSHFS (if required).As a next step you need to submit the SSH key use these links as:
As of February 2020, SSH key submission not accepted via email anymore. Instead, use the process outline here.
For any help, please contact helpdesk@charite.de (as this site is maintained by Charite GB IT).
"},{"location":"connecting/submit-key/charite/#charite-zugangsportal","title":"Charite Zugangsportal","text":"Key are submitted in the Charite Zugangsportal. As of Feb 4, you have to use the \"test\" version for this.
Go to zugang.charite.de and login.
Follow through the login page until you reach the main menu (it's tedious but we belive in you ;) Click the \"SSH Keys\" button.
Paste your SSH key (starting with ssh-rsa) and ending with the label (usually your email, e.g., john.doe@charite.de) into the box (1) and press append (2). By default, the key can be found in the file ~/.ssh/id_rsa.pub in Linux. If you generated the key in Windows, please paste the copied key from the text box. Repeat as necessary. Optionally, go back to the main menu (3) when done.
If you have generated your SSH key with PuTTy, you must right click on the ppk-file, then choose \"Edit with PuTTYgen\" in the right click menu. Enter your passphrase. Then copy the SSH key out of the upper box (already highlighted in blue).
Check if the key has been added
After you clicked append, your key will be printed back to you (as shown in the blurred picture above).
If your key is not printed back to you then adding the SSH key to zugang.charite.de was not successful. In this case please contact helpdesk@charite.de for assistance as they (Charite GB IT) maintains that system and it is out of our (BIH HPC IT) control.
Once your key has been added, it will take a few minutes for the changes to go live.
"},{"location":"connecting/submit-key/mdc/","title":"Submitting an SSH Key to MDC","text":"For MDC users, SSH keys are submitted through the MDC PersDB interface (see below). PersDB is not maintained by BIH HPC IT but by MDC IT.
Warning
The SSH keys are only activated over night (but automatically). This is out of our control. Contact helpdesk@mdc-berlin.de for more information.
"},{"location":"connecting/submit-key/mdc/#detour-using-mdc-vmware-view-to-get-into-mdc-intranet","title":"Detour: Using MDC VMWare View to get into MDC Intranet","text":"In case you are not within the MDC network, connect to MDC VMWare view first and use the web brower in the Window session.
~/.ssh/id_rsa.pub into the clipboard window. Ensure that the whole file contents is there (should end with your email address). If you generated the key in Windows, please paste the copied key from the text box.Thus, you will only be able to connect the next day. - Bask in the glory of having completed this process.
"},{"location":"cubit/","title":"Overview","text":"The static data installation can be found at /data/cephfs-1/work/projects/cubit/18.12/static_data.
The static data directory contains a sub-directory for the genomes, the precomputed index files for several different popular mapping tools and associated annotation (GFF and GTF files) from Ensembl and GENCODE for each of the available genomes. The top-level directory structure is as follows:
static_data/annotationsapp_supportdbexome_panelexon_listprecomputedreferenceThe following Ensembl and GENCODE versions corresponding to the indicated reference genomes will be made available on the cluster.
Database Version Reference Genome Ensembl 65 NCBIM37 (Ensembl release corresponding to GENCODE M1) Ensembl 67 NCBIM37 (Ensembl release for sanger mouse genome assembly) Ensembl 68 GRCm38 (Ensembl release for sanger mouse genome assembly) Ensembl 74 GRCh37 (Ensembl release for GENCODE 19) Ensembl 75 GRCh37 (Latest release for GRCh37) Ensembl 79 GRCh38 (Ensembl release for GENCODE 22) Ensembl 80 GRCh38 (Ensembl release corresponding to GENCODE 22) Ensembl 80 GRCm38 (Ensembl release corresponding to GENCODE M1) GENCODE M1 NCBIM37 (No gff3 file) GENCODE M5 GRCm38 GENCODE 19 current for GRCh37 GENCODE 22 current for GRCh38The annotation files associated with the indicated genomes can be accessed in the following directories:
static_data/annotation\n\u251c\u2500\u2500 ENSEMBL\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 65\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 67\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 68\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCm38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 74\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 75\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 79\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 80\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCm38\n\u2514\u2500\u2500 GENCODE\n \u251c\u2500\u2500 19\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n \u251c\u2500\u2500 22\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh38\n \u251c\u2500\u2500 M1\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n \u2514\u2500\u2500 M5\n \u2514\u2500\u2500 GRCm38\nThe static_data/app_support directory contains all data files that are shipped with a software package installed in cubit. For blast this is not complete and more databases can be added upon request.
static_data/app_support\n\u251c\u2500\u2500 blast\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 variable\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 nt\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 refseq_protein\n\u251c\u2500\u2500 Delly\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.6.5\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.6.7\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.7.1\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.7.2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.7.3\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 0.7.5\n\u251c\u2500\u2500 GATK_bundle\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 2.8\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 b37\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 hg19\n\u251c\u2500\u2500 Jannovar\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.14\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 0.16\n\u251c\u2500\u2500 kraken\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 0.10.5-cubi20160426\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 bacvir\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 minikraken_20141208\n\u251c\u2500\u2500 Oncotator\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 v1_ds_Jan262015\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 1000genome_db\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 achilles\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cancer_gene_census\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 ccle_by_gene\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 ccle_by_gp\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 clinvar\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cosmic\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cosmic_fusion\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cosmic_tissue\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 dbNSFP_ds\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 dbsnp\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 dna_repair_genes\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 esp6500SI_v2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 esp6500SI_v2_coverage\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 familial\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 gencode_out2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 gencode_xrefseq\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hgnc\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mutsig\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 oreganno\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 override_lists\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 ref_hg\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 simple_uniprot\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 so_terms\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 tcgascape\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 tumorscape\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 uniprot_aa_annotation\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 uniprot_aa_xform\n\u2514\u2500\u2500 SnpEff\n \u2514\u2500\u2500 4.1\n \u2514\u2500\u2500 data\n \u251c\u2500\u2500 GRCh37.75\n \u251c\u2500\u2500 GRCh38.79\n \u251c\u2500\u2500 GRCm38.79\n \u251c\u2500\u2500 hg19\n \u251c\u2500\u2500 hg38\n \u2514\u2500\u2500 mm10\nThe file formats in the static_data/db folder are mostly .vcf or .bed files. We provide the following databases:
The directory structure is as follows:
static_data/db\n\u251c\u2500\u2500 COSMIC\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 v72\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n\u251c\u2500\u2500 dbNSFP\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 2.9\n\u251c\u2500\u2500 dbSNP\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 b128\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 mm9\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 b142\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 b144\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 b147\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh38\n\u251c\u2500\u2500 DGV\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 2015-07-23\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n\u251c\u2500\u2500 ExAC\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 release0.3\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 release0.3.1\n\u251c\u2500\u2500 giab\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 NA12878_HG001\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 NISTv2.19\n\u251c\u2500\u2500 goldenpath\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 variable\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mm9\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n\u251c\u2500\u2500 SangerMouseGenomesProject\n\u2502 \u2514\u2500\u2500 REL-1211-SNPs_Indels\n\u2502 \u251c\u2500\u2500 mm9\n\u2502 \u2514\u2500\u2500 NCBIM37\n\u2514\u2500\u2500 UK10K_cohort\n \u2514\u2500\u2500 REL-2012-06-02\nThese exome panel data are proprietary and downloaded after registration. In case you want to use them, be sure you have access to them by creating an account at Agilent or Roche to not run into legal trouble.
static_data/exome_panel\n\u251c\u2500\u2500 Agilent\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 SureSelect_Human_All_Exon_V4\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 SureSelect_Human_All_Exon_V5\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 SureSelect_Human_All_Exon_V6\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 SureSelect_Mouse_All_Exon_V1\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mm9\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n\u2514\u2500\u2500 Roche\n \u2514\u2500\u2500 SeqCap_EZ_MedExome\n \u2514\u2500\u2500 GRCh37\nHere we provide exon lists for some human genome assemblies in the .bed-file format. Each file exists with the original coordinates contained and as a version with 10 bp padded on each site (suffix: _plus_10bp.bed). The folder structure is self-explanatory:
static_data/exon_list\n\u251c\u2500\u2500 CCDS\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 15\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 18\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 hg38\n\u2514\u2500\u2500 ENSEMBL\n \u251c\u2500\u2500 74\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n \u2514\u2500\u2500 75\n \u2514\u2500\u2500 GRCh37\nIndex files for
have been precomputed. The index corresponding to each genome is stored in the following directory structure with the above mentioned reference genomes as subfolders (listed here only for Bowtie/1.1.2, same subfolders for the remaining programs):
static_data/precomputed\n\u251c\u2500\u2500 Bowtie\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 1.1.2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 danRer10\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 dm6\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 ecoli\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCm38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hg18\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hg38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mm10\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mm9\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 NCBIM37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 phix\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 sacCer3\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 UniVec\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 UniVec_Core\n\u251c\u2500\u2500 Bowtie2\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 2.2.5\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 [see Bowtie/1.1.2]\n\u251c\u2500\u2500 BWA\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.7.12\n\u2502\u00a0\u00a0 \u2502 \u2514\u2500\u2500 [see Bowtie/1.1.2]\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 0.7.15\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 [see Bowtie/1.1.2]\n\u2514\u2500\u2500 STAR\n \u2514\u2500\u2500 2.4.1d\n \u2514\u2500\u2500 default\n \u00a0\u00a0 \u2514\u2500\u2500 [see Bowtie/1.1.2]\nWe provide the NCBI mouse reference assembly used by the Sanger Mouse Genomics group for NCBIM37 and GRCm38. This is a reliable source where the appropriate contigs have already been selected by experts. NCBIM37 is annotated with Ensembl release 67 and GRCm38 with Ensembl release 68.
"},{"location":"cubit/references/#ucsc-mouse-reference-genome-assemblies","title":"UCSC mouse reference genome assemblies","text":"The assembly sequence is in one file per chromosome and is available for mm9 and mm10. We concatenated all the chromosome files to one final fasta file for each genome assembly.
"},{"location":"cubit/references/#ncbi-human-reference-genome-assemblies","title":"NCBI human reference genome assemblies","text":"The assembly sequence is in one file per chromosome is available for hg18, hg19 and hg38. We concatenated all the chromosome files to one final fasta file for each genome assembly. Additionally, in the subfolder chromosomes we keep the chromosome fasta files separately for hg18 and hg19.
The following directory structure indicates the available genomes. Where there isn't a name for the data set, either the source (e.g. sanger - from the Sanger Mouse Genomes project) or the download date is used to name the sub-directory.
static_data/reference\n\u251c\u2500\u2500 danRer10\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 dm6\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 ecoli\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 GCA_000005845.2_ASM584v2\n\u251c\u2500\u2500 genomemedley\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 1\n\u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 g1k_phase1\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 g1k_phase2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hs37\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 hs37d5\n\u251c\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hs38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hs38a\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 hs38DH\n\u251c\u2500\u2500 GRCm38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 sanger\n\u251c\u2500\u2500 hg18\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 hg38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 mm10\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 mm9\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 NCBIM37\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 sanger\n\u251c\u2500\u2500 phix\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 illumina\n\u251c\u2500\u2500 sacCer3\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 UniVec\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 9\n\u2514\u2500\u2500 UniVec_Core\n \u2514\u2500\u2500 9\nPlease see the section Connection Problems.
"},{"location":"help/faq/#connecting-to-the-cluster-takes-a-long-time","title":"Connecting to the cluster takes a long time.","text":"The most probable cause for this is a conda installation which loads files on login. To disable this behaviour we can put the conda intialisation code behind a bash alias to run it manually later:
In your ~/.bashrc find the conda block:
# >>> conda initialize >>>\n# !! Contents within this block are managed by 'conda init' !!\n...\n# <<< conda initialize <<<\nEncapsulate the entire section in a bash function like this:
conda_init() {\n # >>> conda initialize >>>\n # !! Contents within this block are managed by 'conda init' !!\n ...\n # <<< conda initialize <<<\n}\nFrom now on to use conda you must first run conda_init which then loads the necessary files.
You can also run the bash shell in verbose mode to find out exactly which command is slowing down login:
$ ssh user@hpc-login-1.cubi.bihealth.org bash -iv\nAdministrativa
Request for both systems are handled separately, depending on the user's affiliation with research/service groups.
Hardware and Systems
Bioinformatics Software
packet_write_wait: Connection to XXX : Broken pipe\". How can I fix this?","text":"Try to put the following line at the top of your ~/.ssh/config.
ServerAliveInterval 30\nThis will make ssh send an empty network package to the server. This will prevent network hardware from thinking your connection is unused/broken and terminating it.
If the problem persists, please report it to hpc-helpdesk@bih-charite.de.
"},{"location":"help/faq/#my-job-terminated-before-being-done-what-happened","title":"My job terminated before being done. What happened?","text":"First of all, look into your job logs. In the case that the job was terminated by Slurm (e.g., because it ran too long), you will find a message like this at the bottom. Please look at the end of the last line in your log file.
slurmstepd: error: *** JOB <your job id> ON med0xxx CANCELLED AT 2020-09-02T21:01:12 DUE TO TIME LIMIT ***\nThis indicates that you need to need to adjust the --time limit to your sbatch command.
slurmstepd: error: Detected 2 oom-kill event(s) in step <your job id>.batch cgroup.\nSome of your processes may have been killed by the cgroup out-of-memory handler\nThis indicates that your job tries to use more memory than has been allocated to it. Also see Slurm Scheduler: Memory Allocation
Otherwise, you can use sacct -j JOBID to read the information that the job accounting system has recorded for your job. A job that was canceled (indicated by CANCELED) by the Slurm job scheduler looks like this (ignore the COMPLETED step that is just some post-job step added by Slurm automatically).
# sacct -j _JOBID_\n JobID JobName Partition Account AllocCPUS State ExitCode\n------------ ---------- ---------- ---------- ---------- ---------- --------\n_JOBID_ snakejob.+ medium hpc-ag-xx+ 4 TIMEOUT 0:0\n_JOBID_.bat+ batch hpc-ag-xx+ 4 CANCELLED 0:15\n_JOBID_.ext+ extern hpc-ag-xx+ 4 COMPLETED 0:0\nUse the --long flag to see all fields (and probably pipe it into less as: sacct -j JOBID --long | less -S). Things to look out for:
MaxRSS)?Elapsed)?Note that --long does not show all fields. For example, the following tells us that the given job was above its elapsed time which caused it to be killed.
# sacct -j _JOBID_ --format Timelimit,Elapsed\n Timelimit Elapsed\n---------- ----------\n 01:00:00 01:00:12\n 01:00:13\n 01:00:12\nUse man sacct, sacct --helpformat, or see the Slurm Documentation for options for the --format field of sacct.
This is most probably caused by your job being allocated insufficient memory. Please see the memory part of the answer to My job terminated before being done. What happened?
"},{"location":"help/faq/#how-can-i-create-a-new-project","title":"How can I create a new project?","text":"You can create a project if you are either a group leader of an AG or a delegate of an AG. If this is the case, please follow these instructions.
"},{"location":"help/faq/#i-cannot-create-pngs-in-r","title":"I cannot create PNGs in R","text":"For using the png method, you need to have an X11 session running. This might be the case if you logged into a cluster node using srun --x11 if configured correctly but is not the case if you submitted a bash job. The solution is to use xvfb-run (xvfb = X11 virtual frame-buffer).
Here is the content of an example script:
$ cat img.R\n#!/usr/bin/env Rscript\n\npng('cars.png')\ncars <- c(1, 3, 6, 4, 9)\nplot(cars)\ndev.off()\nHere, it fails without X11:
$ ./img.R\nError in .External2(C_X11, paste(\"png::\", filename, sep = \"\"), g$width, :\n unable to start device PNG\nCalls: png\nIn addition: Warning message:\nIn png(\"cars.png\") : unable to open connection to X11 display ''\nExecution halted\nHere, it works with xvfb-run:
$ xvfb-run ./img.R\nnull device\n 1\n$ ls\ncars.png foo.png img.R Rplots.pdf\nYou can use scontrol show job JOBID to get the details displayed about your jobs. In the example below, we can see that the job is in the PENDING state. The Reason field tells us that the job did not scheduled because the specified dependency was neverfulfilled. You can find a list of all job reason codes in the Slurm squeue documentation.
JobId=863089 JobName=pipeline_job.sh\n UserId=holtgrem_c(100131) GroupId=hpc-ag-cubi(5272) MCS_label=N/A\n Priority=1 Nice=0 Account=(null) QOS=normal\n JobState=PENDING Reason=DependencyNeverSatisfied Dependency=afterok:863087(failed)\n Requeue=1 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0\n RunTime=00:00:00 TimeLimit=08:00:00 TimeMin=N/A\n SubmitTime=2020-05-03T18:57:34 EligibleTime=Unknown\n AccrueTime=Unknown\n StartTime=Unknown EndTime=Unknown Deadline=N/A\n SuspendTime=None SecsPreSuspend=0 LastSchedEval=2020-05-03T18:57:34\n Partition=debug AllocNode:Sid=hpc-login-1:28797\n ReqNodeList=(null) ExcNodeList=(null)\n NodeList=(null)\n NumNodes=1 NumCPUs=1 NumTasks=1 CPUs/Task=1 ReqB:S:C:T=0:0:*:*\n TRES=cpu=1,node=1,billing=1\n Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*\n MinCPUsNode=1 MinMemoryNode=0 MinTmpDiskNode=0\n Features=(null) DelayBoot=00:00:00\n OverSubscribe=OK Contiguous=0 Licenses=(null) Network=(null)\n Command=/data/cephfs-1/work/projects/medgen_genomes/2019-06-05_genomes_reboot/GRCh37/wgs_cnv_export/pipeline_job.sh\n WorkDir=/data/cephfs-1/work/projects/medgen_genomes/2019-06-05_genomes_reboot/GRCh37/wgs_cnv_export\n StdErr=/data/cephfs-1/work/projects/medgen_genomes/2019-06-05_genomes_reboot/GRCh37/wgs_cnv_export/slurm-863089.out\n StdIn=/dev/null\n StdOut=/data/cephfs-1/work/projects/medgen_genomes/2019-06-05_genomes_reboot/GRCh37/wgs_cnv_export/slurm-863089.out\n Power=\n MailUser=(null) MailType=NONE\nIf you see a Reason=ReqNodeNotAvail,_Reserved_for_maintenance then also see Reservations / Maintenances.
For GPU jobs also see \"My GPU jobs don't get scheduled\".
"},{"location":"help/faq/#my-gpu-jobs-dont-get-scheduled","title":"My GPU jobs don't get scheduled","text":"There are only four GPU machines in the cluster (with four GPUs each, hpc-gpu-1 to hpc-gpu-4). Please inspect first the number of running jobs with GPU resource requests:
hpc-login-1:~$ squeue -o \"%.10i %20j %.2t %.5D %.4C %.10m %.16R %.13b\" \"$@\" | grep hpc-gpu- | sort -k7,7\n 1902163 ONT-basecalling R 1 2 8G hpc-gpu-1 gpu:tesla:2\n 1902167 ONT-basecalling R 1 2 8G hpc-gpu-1 gpu:tesla:2\n 1902164 ONT-basecalling R 1 2 8G hpc-gpu-2 gpu:tesla:2\n 1902166 ONT-basecalling R 1 2 8G hpc-gpu-2 gpu:tesla:2\n 1902162 ONT-basecalling R 1 2 8G hpc-gpu-3 gpu:tesla:2\n 1902165 ONT-basecalling R 1 2 8G hpc-gpu-3 gpu:tesla:2\n 1785264 bash R 1 1 1G hpc-gpu-4 gpu:tesla:2\nThis indicates that there are two free GPUs on hpc-gpu-4.
Second, inspect the node states:
hpc-login-1:~$ sinfo -n hpc-gpu-[1-4]\nPARTITION AVAIL TIMELIMIT NODES STATE NODELIST\ndebug* up 8:00:00 0 n/a\nmedium up 7-00:00:00 0 n/a\nlong up 28-00:00:0 0 n/a\ncritical up 7-00:00:00 0 n/a\nhighmem up 14-00:00:0 0 n/a\ngpu up 14-00:00:0 1 drng hpc-gpu-4\ngpu up 14-00:00:0 3 mix med[0301-0303]\nmpi up 14-00:00:0 0 n/a\nThis tells you that hpc-gpu-1 to hpc-gpu-3 have jobs running (\"mix\" indicates that there are free resources, but these are only CPU cores not GPUs). hpc-gpu-4 is shown to be in \"draining state\". Let's look what's going on there.
hpc-login-1:~$ scontrol show node hpc-gpu-4\nNodeName=hpc-gpu-4 Arch=x86_64 CoresPerSocket=16\n CPUAlloc=2 CPUTot=64 CPULoad=1.44\n AvailableFeatures=skylake\n ActiveFeatures=skylake\n Gres=gpu:tesla:4(S:0-1)\n NodeAddr=hpc-gpu-4 NodeHostName=hpc-gpu-4 Version=20.02.0\n OS=Linux 3.10.0-1127.13.1.el7.x86_64 #1 SMP Tue Jun 23 15:46:38 UTC 2020\n RealMemory=385215 AllocMem=1024 FreeMem=347881 Sockets=2 Boards=1\n State=MIXED+DRAIN ThreadsPerCore=2 TmpDisk=0 Weight=1 Owner=N/A MCS_label=N/A\n Partitions=gpu\n BootTime=2020-06-30T20:33:36 SlurmdStartTime=2020-07-01T09:31:51\n CfgTRES=cpu=64,mem=385215M,billing=64\n AllocTRES=cpu=2,mem=1G\n CapWatts=n/a\n CurrentWatts=0 AveWatts=0\n ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s\n Reason=deep power-off required for PSU [root@2020-07-17T13:21:02]\nThe \"State\" attribute indicates the node has jobs running but is currenlty being \"drained\" (accepts no new jobs). The \"Reason\" gives that it has been scheduled for power-off for maintenance of the power supply unit.
"},{"location":"help/faq/#when-will-my-job-be-scheduled","title":"When will my job be scheduled?","text":"You can use the scontrol show job JOBID command to inspect the scheduling information for your job. For example, the following job is scheduled to start at 2022-09-19T07:53:29 (StartTime) and will be terminated if it does not stop before 2022-09-19T15:53:29 (EndTime) For further information, it has been submitted at 2022-09-15T12:24:57 (SubmitTime) and has been last considered by the scheduler at 2022-09-19T07:53:15 (LastSchedEval).
# scontrol show job 4225062\nJobId=4225062 JobName=C2371_2\n UserId=user_c(133196) GroupId=hpc-ag-group(1030014) MCS_label=N/A\n Priority=805 Nice=0 Account=hpc-ag-group QOS=normal\n JobState=PENDING Reason=QOSMaxCpuPerUserLimit Dependency=(null)\n Requeue=1 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0\n RunTime=00:00:00 TimeLimit=08:00:00 TimeMin=N/A\n SubmitTime=2022-09-15T12:24:57 EligibleTime=2022-09-15T12:24:57\n AccrueTime=2022-09-15T12:24:57\n StartTime=2022-09-19T07:53:29 EndTime=2022-09-19T15:53:29 Deadline=N/A\n SuspendTime=None SecsPreSuspend=0 LastSchedEval=2022-09-19T07:53:15 Scheduler=Main\n Partition=medium AllocNode:Sid=hpc-login-1:557796\n ReqNodeList=(null) ExcNodeList=(null)\n NodeList=(null)\n NumNodes=1-1 NumCPUs=25 NumTasks=25 CPUs/Task=1 ReqB:S:C:T=0:0:*:*\n TRES=cpu=25,mem=150G,node=1,billing=25\n Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*\n MinCPUsNode=1 MinMemoryNode=150G MinTmpDiskNode=0\n Features=(null) DelayBoot=00:00:00\n OverSubscribe=YES Contiguous=0 Licenses=(null) Network=(null)\n Command=/data/cephfs-1/home/users/user_c/work/SCZ_replic/JR_sims/GS_wrapy/wrap_y0_VP_2371_GS_chunk2_C02.sh\n WorkDir=/data/cephfs-1/home/users/user_c/work/SCZ_replic/JR_sims\n StdErr=/data/cephfs-1/home/users/user_c/work/SCZ_replic/JR_sims/E2371_2.txt\n StdIn=/dev/null\n StdOut=/data/cephfs-1/home/users/user_c/work/SCZ_replic/JR_sims/slurm-4225062.out\n Power=\nYou can see the partition that your job runs in with squeue -j JOBID:
hpc-login-1:~$ squeue -j 877092\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 877092 medium snakejob holtgrem R 0:05 1 med0626\nSee Job Scheduler for information about the partition's properties and how jbos are routed to partitions. You can force jobs to run in a particular partition by specifying the --partition parameter, e.g., by adding --partition=medium or -p medium to your srun and sbatch calls.
This is probably answered by the answer to My jobs don't run in the partition I expect.
"},{"location":"help/faq/#how-can-i-mount-a-network-volume-from-elsewhere-on-the-cluster","title":"How can I mount a network volume from elsewhere on the cluster?","text":"You cannot.
"},{"location":"help/faq/#how-can-i-make-workstationserver-files-available-to-the-hpc","title":"How can I make workstation/server files available to the HPC?","text":"You can transfer files to the cluster through Rsync over SSH or through SFTP to the hpc-transfer-1 or hpc-transfer-2 node.
Do not transfer files through the login nodes. Large file transfers through the login nodes can cause performance degradation for the users with interactive SSH connections.
"},{"location":"help/faq/#how-can-i-circumvent-invalid-instruction-signal-4-errors","title":"How can I circumvent \"invalid instruction\" (signal 4) errors?","text":"Make sure that software is compiled with \"sandy bridge\" optimizations and no later one. E.g., use the -march=sandybridge argument to the GCC/LLVM compiler executables.
If you absolutely need it, there are some boxes with more recent processors in the cluster (e.g., Haswell architecture). Look at the /proc/cpuinfo files for details.
Please check whether there might be other jobs waiting in front of you! The following squeue call will show the allocated GPUs of jobs in the gpu queue. This is done by specifying a format string and using the %b field.
squeue -o \"%.10i %9P %20j %10u %.2t %.10M %.6D %10R %b\" -p gpu\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(R TRES_PER_NODE\n 872571 gpu bash user1 R 15:53:25 1 hpc-gpu-3 gpu:tesla:1\n 862261 gpu bash user2 R 2-16:26:59 1 hpc-gpu-4 gpu:tesla:4\n 860771 gpu kidney.job user3 R 2-16:27:12 1 hpc-gpu-2 gpu:tesla:1\n 860772 gpu kidney.job user3 R 2-16:27:12 1 hpc-gpu-2 gpu:tesla:1\n 860773 gpu kidney.job user3 R 2-16:27:12 1 hpc-gpu-2 gpu:tesla:1\n 860770 gpu kidney.job user3 R 4-03:23:08 1 hpc-gpu-1 gpu:tesla:1\n 860766 gpu kidney.job user3 R 4-03:23:11 1 hpc-gpu-3 gpu:tesla:1\n 860767 gpu kidney.job user3 R 4-03:23:11 1 hpc-gpu-1 gpu:tesla:1\n 860768 gpu kidney.job user3 R 4-03:23:11 1 hpc-gpu-1 gpu:tesla:1\nIn the example above, user1 has one job with one GPU running on hpc-gpu-3, user2 has one job running with 4 GPUs on hpc-gpu-4 and user3 has 7 jobs in total running of different machines with one GPU each.
"},{"location":"help/faq/#how-can-i-access-graphical-user-interfaces-such-as-for-matlab-on-the-cluster","title":"How can I access graphical user interfaces (such as for Matlab) on the cluster?","text":"-X for Linux/Mac sshsrun --pty --x11 bash -i (instead of srun --pty --x11 bash -i).Also see:
This is sometimes useful, e.g., for monitoring the CPU/GPU usage of your job interactively.
No Computation Outside of Slurm
Do not perform any computation outside of the scheduler as (1) this breaks the purpose of the scheduling system and (2) administration is not aware and might kill you jobs.
The answer is simple, just SSH into this node.
hpc-login-1:~$ ssh hpc-cpu-xxx\nClassically, jobs on HPC systems are written in a way that they can run on multiple nodes at once, using the network to communicate. Slurm comes from this world and when allocating more than one CPU/core, it might allocate them on different nodes. Please use --nodes=1 to force Slurm to allocate them on a single node.
You can select the CPU architecture by using the -C/--constraint flag to sbatch and srun. The following are available (as detected by the Linux kernel):
ivybridge (96 nodes, plus 4 high-memory nodes)haswell (16 nodes)broadwell (112 nodes)skylake (16 nodes, plus 4 GPU nodes)You can specify contraints with OR such as --constraint=haswell|broadwell|skylake. You can see the assignment of architectures to nodes using the sinfo -o \"%8P %.5a %.10l %.6D %.6t %10f %N\" command. This will also display node partition, availability etc.
No worries!
As documented in the Storage Locations section, each user/project/group has three storage volumes: A small home, a larger work and a large (but temporary) scratch. There are limits on the size of these volumes. You get a nightly warning email in case you are over the soft limit and you will not be able to write any more data if you get above the hard limit. When you login to the login nodes, the quotas and current usage is displayed to you.
Please note that not all files will be displayed when using ls. You have to add the -a parameter to also show files and directory starting with a dot. Often, users are confused if these dot directories take up all of their home quota.
Use the following command to list all files and directories in your home:
hpc-login-1:~$ ls -la ~/\nFor more information on how to keep your home directory clean and avoid quota warnings, please read Home Folder Quota.
"},{"location":"help/faq/#im-getting-a-disk-quota-exceeded-error","title":"I'm getting a \"Disk quota exceeded\" error.","text":"Most probably you are running into the same problem as described above: Help, I'm getting a Quota Warning Email!
"},{"location":"help/faq/#environment-modules-dont-work-and-i-get-module-command-not-found","title":"Environment modules don't work and I get \"module: command not found\"","text":"First of all, ensure that you are on a compute node and not on one of the login nodes. One common reason is that the system-wide Bash configuration has not been loaded, try to execute source /etc/bashrc and then re-try using module. In the case that the problem persists, please contact hpc-helpdesk@bih-charite.de.
All users get their home directory setup using a skelleton files. These file names start with a dot . and are hidden when you type ls, you have to type ls -a to see them. You can find the current skelleton in /etc/skel.bih and inspect the content of the Bash related files as follows:
hpc-login-1:~$ head /etc/skel.bih/.bash*\n==> /etc/skel.bih/.bash_logout <==\n# ~/.bash_logout\n\n==> /etc/skel.bih/.bash_profile <==\n# .bash_profile\n\n# Get the aliases and functions\nif [ -f ~/.bashrc ]; then\n . ~/.bashrc\nfi\n\n# User specific environment and startup programs\n\nPATH=$PATH:$HOME/.local/bin:$HOME/bin\n\n==> /etc/skel.bih/.bashrc <==\n# .bashrc\n\n# Source global definitions\nif [ -f /etc/bashrc ]; then\n . /etc/bashrc\nfi\n\n# Uncomment the following line if you don't like systemctl's auto-paging feature:\n# export SYSTEMD_PAGER=\nThere actually are a couple of more files by default. The original copy in /etc/skel.bih might slightly change over time during improvements but we will not touch your home directory in an unsolicited way at any time!
hpc-login-1:~$ tree -a /etc/skel.bih/\n/etc/skel.bih/\n\u251c\u2500\u2500 .bash_logout\n\u251c\u2500\u2500 .bash_profile\n\u251c\u2500\u2500 .bashrc\n\u251c\u2500\u2500 .screenrc\n\u2514\u2500\u2500 .vimrc\nHave a look at our How-To: Debug Software and How-To: Debug Software on HPC Systems guides!
But it works on my workstation!
Yes, please also refer to these guides on possible approaches to find the problem.
"},{"location":"help/faq/#my-r-sessions-is-always-crashing","title":"My R sessions is always crashing!","text":"We are aware of issues for macOS users that are related to locales. You might see something like the following in your R session or logs.
Note that this is not macOS specific by itself but by macOS using a locale that R does not understand.
Try using export LC_ALL=C before calling R. If the problem goes away, you can add the following lines to your ~/.bashrc file to make this permanent. Note that this is actually a good idea so your locale settings are fixed.
# Fix all locale settings to \"C\" for robustness.\nexport LC_ALL=C\nChatGPT on Unix Locales
Let's se what ChatGPT 3.5 tells us about Unix locale.
Me: Please explain the concept of Unix locales to me.
Certainly! Unix locales are a set of environment variables and settings that define how various aspects of software and user interactions should be handled to accommodate different languages, regions, and cultural conventions. They play a crucial role in enabling software to adapt to local preferences for displaying text, formatting numbers, handling dates, and more.
In a Unix-like operating system (such as Linux or macOS), locales provide a way to customize how programs behave based on the user's language, country, and cultural norms. Here are some of the key components of Unix locales:
Language: The primary language used for messages, menus, prompts, and system output. For instance, if you set your locale to \"en_US\", messages and interfaces will be in English (United States).
Region: The country or region that defines conventions for aspects like date and time formatting, number formatting (decimal point or comma, thousands separator, etc.), and currency symbols. For example, \"en_GB\" represents English (Great Britain).
Character Encoding: Specifies how characters are encoded, which is vital for correctly displaying and handling text in various scripts and languages. Common encodings include UTF-8 for Unicode text.
Collation: Defines how strings should be sorted. Different languages have different sorting orders for characters, and the locale determines the appropriate collation rules.
Numeric and Date Formats: Specifies how numbers, dates, and times should be formatted. Different regions have different conventions for date formats (dd/mm/yyyy vs. mm/dd/yyyy) and numerical representations (commas vs. periods as decimal points).
Currency Symbols: Defines the symbol or abbreviation used for representing currency. This is important for currency formatting in software.
Unix locales are typically identified by a combination of language and region codes, like \"en_US\" for American English or \"fr_FR\" for French (France). These locales are stored in files on the system, and they contain the necessary rules and data for proper localization.
Users and administrators can set the system's default locale, and individual users can override this setting in their environment. Here are some common commands used to manage locales:
locale: Displays the current locale settings.locale -a: Lists all available locales on the system.export LC_ALL=locale_code: Temporarily sets the locale for the current session.update-locale locale_code: Changes the system-wide default locale.Programs that are designed to be locale-aware use these settings to display messages, format numbers and dates, and handle other language-dependent features. This allows software to be more user-friendly and relevant to different linguistic and cultural contexts.
In summary, Unix locales provide a framework for adapting software behavior to match the language, region, and cultural preferences of users, ensuring a more inclusive and globally accessible computing experience.
"},{"location":"help/faq/#which-cuda-version-is-installed","title":"Which CUDA version is installed?","text":"For this, connect to the node you want to query (via SSH but do not perform any computation via SSH!)
hpc-login-1:~$ ssh hpc-gpu-1\nhpc-gpu-1:~$ yum list installed 2>/dev/null | grep cuda.x86_64\ncuda.x86_64 10.2.89-1 @local-cuda\nnvidia-driver-latest-dkms-cuda.x86_64 3:440.64.00-1.el7 @local-cuda\nNo, as Docker essentially gives you access as the root user.
However, you can use Apptainer (former Singularity) to run containers (and even many Docker contains if they are \"properly built\"). Also see Using Apptainer (with Docker Images).
"},{"location":"help/faq/#how-can-i-copy-data-between-the-max-cluster-mdc-network-and-bih-hpc","title":"How can I copy data between the MAX Cluster (MDC Network) and BIH HPC?","text":"The MAX cluster is the HPC system of the MDC. It is located in the MDC network. The BIH HPC is located in the BIH network.
In general, connections can only be initiated from the MDC network to the BIH network. The reverse does not work. In other words, you have to log into the MAX cluster and then initiate your file copies to or from the BIH HPC from there. E.g., use rsync -avP some/path user_m@hpc-transfer-1.cubi.bihealth.org:/another/path to copy files from the MAX cluster to BIH HPC and rsync -avP user_m@hpc-transfer-1.cubi.bihealth.org:/another/path some/path to copy data from the BIH HPC to the MAX cluster.
In general, connections can only be initiated from the Charit\u00e9 network to the BIH network. The reverse does not work. In other words, you have to be on a machine inside the Charit\u00e9 network and then initiate your file copies to or from the BIH HPC from there. E.g., use rsync -avP some/path user_c@hpc-transfer-1.cubi.bihealth.org:/another/path to copy files from the MAX cluster to BIH HPC and rsync -avP user_c@hpc-transfer-1.cubi.bihealth.org:/another/path some/path to copy data from the BIH HPC to the MAX cluster.
As of December 3, 2020 we have established a policy to limit you to 512 files and 128MB of RAM. Further, you are limited to using the equivalent of one core. This limit is enforced for all processes originating from an SSH session and the limit is enforced on all jobs. This was done to prevent users from thrashing the head nodes or using SSH based sessions for computation.
"},{"location":"help/faq/#slurm-complains-about-execve-no-such-file-or-directory","title":"Slurm complains aboutexecve / \"No such file or directory\"","text":"This means that the program that you want to execute does not exist. Consider the following example:
[user@hpc-login-1 ~]$ srun --time 2-0 --nodes=1 --ntasks-per-node=1 \\\n --cpus-per-task=12 --mem 96G --partition staging --immediate 5 \\\n --pty bash -i\nslurmstepd: error: execve(): 5: No such file or directory\nsrun: error: hpc-cpu-2: task 0: Exited with exit code 2\nCan you spot the problem? In this case, the problem is that for long arguments such as --mem you must use the equal sign for --arg=value with Slurm. This means that instead of writing --mem 96G --partition staging --immediate 5, you must use `--mem=96G --partition=staging --immediate=5.
In this respect, Slurm deviates from the GNU argument syntax where the equal sign is optional for long arguments.
"},{"location":"help/faq/#slurmstepd-says-that-hwloc_get_obj_below_by_type-fails","title":"slurmstepd says that hwloc_get_obj_below_by_type fails","text":"You can ignore the following problem:
slurmstepd: error: hwloc_get_obj_below_by_type() failing, task/affinity plugin may be required to address bug fixed in HWLOC version 1.11.5\nslurmstepd: error: task[0] unable to set taskset '0x0'\nThis is a minor failure related to Slurm and cgroups. Your job should run through successfully despite this error (that is more of a warning for end-users).
"},{"location":"help/faq/#how-can-i-share-filescollaborate-with-users-from-another-work-group","title":"How can I share files/collaborate with users from another work group?","text":"Please use projects as documented here. Projects were created for this particular purpose.
"},{"location":"help/faq/#whats-the-relation-of-charite-mdc-and-cluster-accounts","title":"What's the relation of Charite, MDC, and cluster accounts?","text":"For HPC 4 Research either an active and working Charite or MDC account is required (that is, you can login e.g., into email.charite.de or mail.mdc-berlin.de). The system has a separate meta directory that is used for the authorization of users (in other words, whether the user is active, has access to the system, and which groups the user belongs to). Charite and MDC accounts map to accounts <Charite user name>_c and <MDC user name>_m accounts in this meta directory. In the case that a user has both Charite and MDC accounts these are completely separate entities in the meta directory. For authentication (veryfing that a user has acccess to an account), the Charite and MDC account systems (MS Active Directory) are used. Authentication currently only uses the SSH keys deposited into Charite (via zugang.charite.de) and MDC (via MDC persdb). Users have to obtain a suitable Charite/MDC account via Charite and MDC central IT departments and upload their SSH keys through the host organization systems on their own. The hpc-helpdesk process is then used for getting their accounts setup on the HPC 4 Research system (the home/work/scratch shares being setup), becoming part of the special hpc-users group that controls access to the system and organizing users into work groups and projects.
The process of submitting keys to Charite and MDC is documented in the \"Connecting\" section.
"},{"location":"help/faq/#how-do-charitemdccluster-accounts-interplay-with-vpn-and-the-mdc-jail-node","title":"How do Charite/MDC/Cluster accounts interplay with VPN and the MDC jail node?","text":"Charite users have to obtain a VPN account with the appropriate VPN access permissions, i.e., Zusatzantrag B as documented here. For Charite VPN, as for all Charite IT systems, users must use their Charite user name (e.g., jdoe and not jdoe_c).
MDC users either have to use MDC VPN or the MDC jail node, as documented here. For MDC VPN and jail node, as for all MDC IT systems, users must use their MDC user name (e.g., jdoe and not jdoe_m).
For help with VPN or jail node, please contact the central Charite or MDC helpdesks as appropriate.
Only when connecting from the host organizations' VPN or from the host organizations' jail node, the users use the HPC 4 Research user name that is jdoe_c or jdoe_m and not jdoe!
BIH HPC IT does not have the resources to offer such a service to normal users.
In particular, for privacy sensitive data this comes with a large number of strings attached to fulfill all regulatory requirements. If you need to exchange such data then you need to contact the central IT departments of your home organisation:
If your data is not privacy sensitive or you can guarantee strong encryption of the data then the Gigamove service of RWTH Aachen might come in handy:
You can login via Charite/MDC credentials (or most German academic institutions) and store up to 1TB of data at a time in the account with each file having up to 100GB.
As a note, Charite GB IT has a (German) manual on how to use 7-Zip with AES256 and strong passwords for encrypting data such that it is fit for transfer over unencrypted channels. You can find it here (Charite Intranet only) at point 2.12.
The key point is using a strong password (e.g. with the pwgen utility), creating an encrypted file with AES256 encryption, using distinct password for each recipient, and exchanging the password over a second channel (SMS or voice phone). Note that the central manual remains the ground truth of information and this FAQ entry may not reflect the current process recommended by GB IT if it changes without us noticing.
Can you solve the question yourself?
Please try to solve the question yourself with this manual and Google.
If the problem turns out to be hard, we're happy to help.
This page describes how to write a good help request ticket.
There is more specific questions for common issues given below.
"},{"location":"help/good-tickets/#problems-connecting-to-the-cluster","title":"Problems Connecting to the Cluster","text":"ifconfig on Linux/Mac, ipconfig on Windows)?ssh-add -l and add -vvv to the SSH command that fails for you.scontrol show job <jobid> or sacct --long -j <jobid> of your job.Getting Help
Our helpdesk can be reached via email to hpc-helpdesk@bih-charite.de. Please read our guide on how to write good tickets first.
Please also use the handy figure below on general problem resolution.
But before contacting the helpdesk, try to get help in the HPC Talk BIH HPC user self-help forum!
"},{"location":"help/helpdesk/#helpdesk-scope","title":"Helpdesk Scope","text":"Our helpdesk can support you in the following areas:
We will try our best to resolve these issues. Please note that all other questions can only be answered in a \"best effort way\".
"},{"location":"help/helpdesk/#helpdesk-non-scope","title":"Helpdesk Non-Scope","text":"The following topics are out of scope for the BIH HPC Helpdesk:
We're happy to see if we can help when there is a concrete problem with the software, e.g.,
Another community-driven possibility to get help is our \u201cHPC Talk\u201d forum. After this manual, it should be the first place to consult.
https://hpc-talk.cubi.bihealth.org/
Its main purpose is to serve as a FAQ, so with time and more people participating, you will more likely find an answer to your question. We also use it to make announcements and give an up-to-date status of current problems with the cluster, so it is worth logging in every once in a while. It is also a great first place to look at if you're experiencing problems with the cluster. Maybe it's a known issue.
Despite users also being able to answer questions, our admins do participate on a regular basis.
"},{"location":"how-to/connect/gpu-nodes/","title":"How-To: Connect to GPU Nodes","text":"The cluster has seven nodes with four Tesla V100 GPUs each: hpc-gpu-{1..7} and one node with 10 A40 GPUs: hpc-gpu-8.
Connecting to a node with GPUs is easy. You request one or more GPU cores by adding a generic resources flag to your Slurm job submission via srun or sbatch.
--gres=gpu:tesla:COUNT will request NVIDIA V100 cores.--gres=gpu:a40:COUNT will request NVIDIA A40 cores.--gres=gpu:COUNT will request any available GPU cores.Your job will be automatically placed in the Slurm gpu partition and allocated a number of COUNT GPUs.
Info
Fair use rules apply. As GPU nodes are a limited resource, excessive use by single users is prohibited and can lead to mitigating actions. Be nice and cooperative with other users. Tip: getent passwd USER_NAME will give you a user's contact details.
Interactive Use of GPU Nodes is Discouraged
While interactive computation on the GPU nodes is convenient, it makes it very easy to forget a job after your computation is complete and let it run idle. While your job is allocated, it blocks the allocated GPUs and other users cannot use them although you might not be actually using them. Please prefer batch jobs for your GPU jobs over interactive jobs.
Furthermore, interactive GPU jobs are currently limited to 24 hours. We will monitor the situation and adjust that limit to optimize GPU usage and usability.
Please also note that allocation of GPUs through Slurm is mandatory, in other words: Using GPUs via SSH sessions is prohibited. The scheduler is not aware of manually allocated GPUs and this interferes with other users' jobs.
"},{"location":"how-to/connect/gpu-nodes/#usage-example","title":"Usage example","text":""},{"location":"how-to/connect/gpu-nodes/#preparation","title":"Preparation","text":"We will setup a miniforge installation with pytorch testing the GPU. If you already have this setup then you can skip this step
hpc-login-1:~$ srun --pty bash\nhpc-cpu-1:~$ wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-x86_64.sh\nhpc-cpu-1:~$ bash Miniforge3-Linux-x86_64.sh -b -p ~/work/miniforge\nhpc-cpu-1:~$ source ~/work/miniforge/bin/activate\nhpc-cpu-1:~$ conda create -y -n gpu-test pytorch cudatoolkit=10.2 -c pytorch\nhpc-cpu-1:~$ conda activate gpu-test\nhpc-cpu-1:~$ python -c 'import torch; print(torch.cuda.is_available())'\nFalse\nhpc-cpu-1:~$ exit\nhpc-login-1:~$\nThe False shows that CUDA is not available on the node but that is to be expected. We're only warming up!
Let us now allocate a GPU. The Slurm schedule will properly allocate GPUs for you and setup the environment variable that tell CUDA which devices are available. The following dry run shows these environment variables (and that they are not available on the login node).
hpc-login-1:~$ export | grep CUDA_VISIBLE_DEVICES\nhpc-login-1:~$ srun --gres=gpu:tesla:1 --pty bash\nhpc-gpu-1:~$ export | grep CUDA_VISIBLE_DEVICES\ndeclare -x CUDA_VISIBLE_DEVICES=\"0\"\nhpc-gpu-1:~$ exit\nhpc-login-1:~$ srun --gres=gpu:tesla:2 --pty bash\nhpc-gpu-1:~$ export | grep CUDA_VISIBLE_DEVICES\ndeclare -x CUDA_VISIBLE_DEVICES=\"0,1\"\nAs you see, you can also reserve multiple GPUs. If we were to open two concurrent connections (e. g. in a screen) to the same node when allocating one GPU each, the allocated GPUs would be non-overlapping. Note that any two jobs are isolated using Linux cgroups (\"container\" technology) so you cannot accidentally use a GPU of another job.
Now to the somewhat boring part where we show that CUDA actually works.
hpc-login-1:~$ srun --gres=gpu:tesla:1 --pty bash\nhpc-gpu-1:~$ nvcc --version\nnvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2019 NVIDIA Corporation\nBuilt on Wed_Oct_23_19:24:38_PDT_2019\nCuda compilation tools, release 10.2, V10.2.89\nhpc-gpu-1:~$ source ~/work/miniforge/bin/activate\nhpc-gpu-1:~$ conda activate gpu-test\nhpc-gpu-1:~$ python -c 'import torch; print(torch.cuda.is_available())'\nTrue\nNote
If scheduling a GPU fails, consider explicitely requesting the GPU partion via --partition gpu (or #SBATCH --partition gpu).
Also make sure to read the FAQ entry \"I have problems connecting to the GPU node! What's wrong?\" if you encounter problems.
"},{"location":"how-to/connect/gpu-nodes/#bonus-1-who-is-using-the-gpus","title":"Bonus #1: Who is using the GPUs?","text":"Use squeue to find out about currently queued jobs (the egrep only keeps the header and entries in the gpu partition).
hpc-login-1:~$ squeue | egrep -iw 'JOBID|gpu'\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 33 gpu bash holtgrem R 2:26 1 hpc-gpu-1\nTo find out how active the GPU nodes actually are, you can connect to the nodes (without allocating a GPU; you can do this even if the node is full) and then use nvidia-smi.
hpc-login-1:~$ ssh hpc-gpu-1 bash\nhpc-gpu-1:~$ nvidia-smi\nFri Mar 6 11:10:08 2020\n+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 440.33.01 Driver Version: 440.33.01 CUDA Version: 10.2 |\n|-------------------------------+----------------------+----------------------+\n| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |\n| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |\n|===============================+======================+======================|\n| 0 Tesla V100-SXM2... Off | 00000000:18:00.0 Off | 0 |\n| N/A 62C P0 246W / 300W | 16604MiB / 32510MiB | 99% Default |\n+-------------------------------+----------------------+----------------------+\n| 1 Tesla V100-SXM2... Off | 00000000:3B:00.0 Off | 0 |\n| N/A 61C P0 270W / 300W | 16604MiB / 32510MiB | 100% Default |\n+-------------------------------+----------------------+----------------------+\n| 2 Tesla V100-SXM2... Off | 00000000:86:00.0 Off | 0 |\n| N/A 39C P0 55W / 300W | 0MiB / 32510MiB | 0% Default |\n+-------------------------------+----------------------+----------------------+\n| 3 Tesla V100-SXM2... Off | 00000000:AF:00.0 Off | 0 |\n| N/A 44C P0 60W / 300W | 0MiB / 32510MiB | 4% Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes: GPU Memory |\n| GPU PID Type Process name Usage |\n|=============================================================================|\n| 0 43461 C python 16593MiB |\n| 1 43373 C python 16593MiB |\n+-----------------------------------------------------------------------------+\nNote that allocating a GPU makes it unavailable for everyone else, so please behave nicely and be cooperative. If you see someone blocking the GPU nodes for a long time, first find out who it is. You can type getent passwd USER_NAME on any cluster node to see their email address (and work phone number if added). Send a friendly email, most likely they blocked the node accidentally. If you cannot resolve the issue (e. g. the user is not reachable) then please contact hpc-helpdesk@bih-charite.de.
The cluster has 4 high-memory nodes with 1.5 TB of RAM. You can connect to these nodes using the highmem SLURM partition (see below). Jobs allocating more than 200 GB of RAM are automatically routed to the highmem nodes.
Info
Fair use rules apply. As high-memory nodes are a limited resource, excessive use by single users is prohibited and can lead to mitigating actions. Be nice and cooperative with other users. Tip: getent passwd USER_NAME will give you a user's contact details.
In the cluster there are four High-memory used which can be used:
hpc-login-1:~$ sinfo -p highmem\nPARTITION AVAIL TIMELIMIT NODES STATE NODELIST \nhighmem up 14-00:00:0 3 idle med040[1-4] \nTo connect to one of them, simply allocate more than 200GB of RAM in your job.
hpc-login-1:~$ srun --pty --mem=300GB bash -i\nmed0401:~$\nYou can also pick one of the hostnames:
hpc-login-1:~$ srun --pty --mem=300GB --nodelist=med0403 bash -i\nmed0403:~$\nAfter successfull login, you can see that you are in \"highmem\" queue:
med0403:~$ squeue\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) \n[...]\n 270 highmem bash holtgrem R 1:25 1 med0403 \nClick on the edit link at the top of each page as shown below.
Please Contribute!
This guide is far from complete. Please feel free to contribute, e.g., refer to How-To: Contribute to this Document.
Please make sure that you have read How-To: Debug Software as a general primer.
As debugging is hard enough already, it makes one wonder how to do this on the HPC system in batch mode. Here is a list of pointers.
"},{"location":"how-to/misc/debug-at-hpc/#attempt-1-run-it-interactively","title":"Attempt 1: Run it interactively!","text":"First of all, you can of course get an interactive session using srun --pty bash -i and then run your program interactively. Make sure to allocate appropriate memory and cores for your purpose. You might also want to first start a screen or tmux session on the login node such that network interruptions to the login node don't harm your hard debugging work!
Does the program work correctly if you do this? If yes, and it only fails when run in batch mode, consider the following behaviour of the scheduler.
The scheduler takes your resource requirements and tries to find a free slot. Once it has found a free slot, it will attempt to run the program. This mainly differs in running it interactively in standard input, output, and error streams.
/dev/null such that no input is read. You can change this with the --input= flag to specify a file.--output=. You can use certain wildcards to make the output (but also the input files) depend on things like the job ID or job name.mkdir it in the job script itself.Please refer to the sbatch documentation for details.
If your program fails without leaving any log file or any other trace, make sure that the path to the output file exists. To the best of the author's knowledge, there is no way to tell apart a crash because this does not exist and a program failure (except maybe for the running time of 0 seconds and memory usage of 0 bytes).
"},{"location":"how-to/misc/debug-at-hpc/#attempt-2-inspect-the-logs","title":"Attempt 2: Inspect the logs","text":"Do you see any exception in your log files? If not, continue.
If your job is canceled by scancel or stopped because it exhausted it maximal running time or allocated resources then you will find a note in the last line of your error output log (usually folded into the standard output). Please note that if the previous output line did not include a line ending, the message might be at the very end of the last line.
The message will look similar to:
slurmstepd: error: *** JOB <your job id> ON med0xxx CANCELLED AT 2020-09-02T21:01:12 DUE TO TIME LIMIT ***\nIdeally, you can add one or more --verbose/-v flags to your program to increase verbosity. See how far your program gets, see where it fails. This attempt will be greatly helped by reproducible running on a minimal working example.
sattach","text":"You can use sattach for attaching your terminal to your running job. This way, you can perform an interactive inspection of the commands.
You can combine this with one of the next attempst of using debuggers to e.g., get an pdb debugger at an important position of your program. However, please note that pdb and ipdb will stop the program's execution if the standard input stream is at end of file (which /dev/null is and this is used by default in sbatch jobs).
Log into the node that your program runs on either using srun --pty --nodelist=NODE or using ssh. Please note that you should never perform computational intensive things when logging into the node directly. You can then use all activity inspection tips from How-To: Debug Software.
After having logged into the node running your program, you can of course also attach to the program with gdb -p PID or cgdb -p PID.
Here are some final remarks:
Please Contribute!
This guide is far from complete. Please feel free to contribute, e.g., refer to How-To: Contribute to this Document.
Software development in general or even debugging of software are very broad topics. As such, we will not be able to handle them here comprehensively. Rather, we will give a tour de force on practical and minimal approaches of debugging of software. Here, debugging refers to the process of locating errors in your program and removing them.
Origin of the term debugging
The terms \"bug\" and \"debugging\" are popularly attributed to Admiral Grace Hopper in the 1940s. While she was working on a Mark II computer at Harvard University, her associates discovered a moth stuck in a relay and thereby impeding operation, whereupon she remarked that they were \"debugging\" the system. However, the term \"bug\", in the sense of \"technical error\", dates back at least to 1878 and Thomas Edison (see software bug for a full discussion).
-- Wikipedia: Debugging
When forgetting a moment about everything known about software engineering, programming roughly work sin the following cycle:
You run your program. In the case of failure, you need to remove the problem until the program runs through. You then start implementing the next change or feature. But how do you actually locate the problem? Let us walk through a couple of steps.
"},{"location":"how-to/misc/debug-software/#step-1-find-out-that-there-is-an-error","title":"Step 1: Find out that there is an error","text":"This might seem trivial but let us think about this for a moment. For this
You could make this step a bit more comfortable by writing a little checker script that compares expected and actual output.
"},{"location":"how-to/misc/debug-software/#step-2-reproduce-your-error","title":"Step 2: Reproduce your error","text":"You will have to find out how often or regularly the problem occurs. Does the problem occur on all inputs or only specific ones? Does it occur with all parameters? Make sure that you can reproduce the problem, otherwise the problem will be hard to track down.
Discard randomness
In most applications, true randomness is neither required nor used in programs. Rather, pseudo random number generators are used that are usually seeded with a special value. In many cases, the current time is used which makes it hard to reproduce problems. Rather, use a fixed seed, e.g., by calling srand(42) in C. You could also make this a parameter of your program, but make sure that you can fix all pseudo randomness in your program so you can deterministically reproduce its behaviour.
Try to find a minimal input set on which you can produce your problem. For example, you could use samtools view FILE.bam chr1:90,000-100,000 to cut out regions from a BAM file. The next step is to nail down the problem. Ideally, you can deactivate or comment out parts of your program that are irrelevant to the problem.
This will allow you to get to the problematic point in your program quicker and make the whole debugging exercise easier on yourself.
"},{"location":"how-to/misc/debug-software/#interlude-what-we-have-up-to-here","title":"Interlude: What we have up to here","text":"We can now
If you reached the points above, you have probably cut the time to resolve the problem by 90% already.
Let us now consider a few things that you can do from here to find the source of your problems.
"},{"location":"how-to/misc/debug-software/#method-1-stare-at-your-source-code","title":"Method 1: Stare at your source code","text":"Again, this is trivial, but: look at your code and try to follow through what it does with your given input. This is nicely complemented with the following methods. ;-)
There is a class of tools to help you in doing this, so-called static code analysis tools. They analyze the source code for problematic patterns. The success and power of such analysis tools tends to corellate strongly with how strictly typed the targeted programming language is. E.g., there are very powerful tools for Java, C/C++. However, there is some useful tool support out there for dynamic languages such as Python.
Here is a short list of pointers to static code analysis tools (feel free to extend the list):
The most simple approach is to use print statements (or similar) to print the current line or value of parameters. While sometimes frowned upon, this certainly is one of the most robust ways to see what is happening in your program. However, beware that too much output might slow down your program or actually make your problem disappear in the case of subtle threading/timing issues (sometimes referred to as \"Heisenbugs\").
Standard output vs. error
Classically, Linux/Unix programs can print back to the user's terminal in two ways: standard output and standard errors. By convention, logging should go to stderr. The standard error stream also has the advantage that writing to it has a more direct effect. In contrast to stdout which is usually setup to be (line) buffered (you will only see output after the next newline character), stderr is unbuffered.
"},{"location":"how-to/misc/debug-software/#look-at-tophtop","title":"Look attop/htop","text":"The tools top and htop are useful tools for inspecting the activity on the current computer. The following parameters are useful (and are actually also available as key strokes when they are running).
-c -- show the programs' command lines-u USER -- show the processes of the userYou can exit either tool by pressing q or Ctrl-C.
Use the man, Luke!
Besides searching the internet for a unix command, you can also read its manual page by running man TOOL. If this does not work, try TOOL --help to see its builtin help function. Also, doing an internet search for \"man tool\" might help.
strace","text":"The program strace allows you to intercept the calls of your program to the kernel. As the kernel is needed for actions such as accessing the network or file system. Thus this is not so useful if your program gets stuck in \"user land\", but this might be useful to see which files it is accessing.
Pro-Tip: if you move the selection line of htop to a process then you can strace the program by pressing s.
lsof","text":"The lsof program lists all open files with the processes that are accessing them. This is useful for seeing which files you program has opened.
You can even build a progress bar with lsof, although that requires sudo privileges which you might not have on the system that you are using.
Pro-Tip: if you move the selection line of htop to a process then you can list the open files by pressing l.
There are more ways of inspecting your program, here are some:
perfLet us now enter the world of interactive debuggers. Integrated development environment (IDEs) generally consist of an editor, a compiler/interpreter, and an ineractive/visual debugger. Usually, they have a debugger program at their core that can also be used on their command line.
"},{"location":"how-to/misc/debug-software/#old-but-gold-gdb","title":"Old but gold:gdb","text":"On Unix systems, a widely used debugger is gdb the GNU debugger. gdb is a command line program and if you are not used to it, it might be hard to use. However, here are some pointers on how to use it:
The commands in interactive mode include:
quit or Ctrl-D to exit the debuggerb file.ext:123 set breakpoint in file.ext on line 123r run the programp var_name print the value of the variable var_namedisplay var_name print the value of the variable var_name every time execution stopsl print the source code around the current line (multiple calls will show the next 10 lines or so, and so on)l 123 print lines around line 123f show information about the current frame (that is the current source location)bt show the backtrace (that is all functions above the current one)n step to the next lines step into function callsfinish run the current function until it returnshelp to get more helpYou can call your program directly with command line arguments using cgdb [cgdb-args] --args path/to/program -- [program-args.You can also attach to running programs usingcgdb -p PIDonce you have found out the process ID to attach to usinghtoporps`.
Pro-tip: use cgdb for an easier to use version that displays the source code in split screen and stores command line histories over sessions.
You can get a simple REPL (read-execute-print loop) at virtually any position in your program by adding:
import pdb; pdb.set_trace()\nYou will get a prompt at the current position and can issue several commands including:
quit or Ctrl-D to exit the debuggerp var_name to print the variable with var_namef show information about the current frame (that is the current source location)bt show the backtrace (that is all functions called above the current one)continue to continue runninghelp to get more helpPro-tip: use import ipdb; ipdb.set_trace() (after installing the ipdb package, of course) to get an IPython-based prompt that is much more comfortable to use.
Here is a free bonus pro-tip: learn how to use version control, e.g., Git. This will allow you to go back to previous versions without problems and see current changes to your source code.
Combine the pro tip on using version control (learn Git already!) with this one: learn how to write automated tests. This will allow you to quickly narrow down problematic changes in your version control history.
Again, testing is another topic alltogether, so here are just some links to testing frameworks to get you started:
The following web resources can serve as a starting point on how to use debuggers.
We provide a user forum using the Discourse software at
First of all, visit the website for the first time: https://hpc-talk.cubi.bihealth.org
You will then be directed to our Single-Sign-On Page.
Use the appropriate button for your host organisation (MDC / Charite) where also your cluster account belongs to.
Then use the usual of your host organisation.
Clicked wrong organisation?
If you accidentally clicked the wrong institution then you need to clear your browser history up to the point where you clicked (e.g., for the last hour).
You will be shown the following screen after the first login.
You can proceed with reading the notification or split it. The site is mostly self-explanatory. let us point you at a couple of interesting things for first steps.
Here you can setup your preferences
Use the \"New Topic\" button to create a new topic. Set a meaningful title, select a suitable category (we will update the list of categories over time), and write down your question or discussion item. Finally, click \"Create Topic\" to create the new topic.
You will be directed to the page with your new topic.
You can enable email notifications to receive emails if someone answers.
"},{"location":"how-to/misc/hpc-talk/#disabling-browser-notifications","title":"Disabling Browser Notifications","text":"In your settings, you will find an option to disable browser notifications in this browser.
Or you can use the do not disturb button.
"},{"location":"how-to/misc/hpc-talk/#closing-remarks","title":"Closing Remarks","text":"We established the HPC Talk forum as a self-help forum for users. Alas, there is a number of such sites out there already that are populated by more users.
How does HPC Talk fit in?
We think it is most useful for asking questions and discussing points that are directly related to the BIH HPC system.
What alternatives do I have?
For example:
Obtaining File Boxes
At the moment, file boxes are only available to members of core facilities (e.g., genomics, bioinformatics, or metabolomics) for exchanging files for their collaboration partners. Currently, HPC users cannot use the file box mechanism on their own.
BIH HPC IT provides a file exchange server to be used by the BIH core facilities and their users. The server is located in the BIH DMZ in Buch. Users authenticate using their Charite/BIH (user@CHARITE) or MDC accounts (user@MDC-BERLIN). File exchange is organized using \"file boxes\", directories created on the server to which selected users are granted access. Access control list maintenance is done with audit-trails (\"Revisionssicherheit\") and the file access itself is also logged to comply with data protection standards.
Access from Charite Network
Access from the Charite network (IP ranges 141.x.x.x and 10.x.x.x) must connect through the Charite proxy (http://proxy.charite.de:8080). Depending on the client software that you are using, you might have to configure the proxy.
File boxes are created by the core facilities (e.g., the genomics facilities at Charite and MDC). The facility members also organize the access control. Please talk to your core facility contact on file exchange.
External users must obtain a Charite or MDC account first. Account creation is handled by the core facilities that the external user is a customer of.
"},{"location":"how-to/service/file-exchange/#file-access","title":"File Access","text":"Generally, you will be given a URL to your file box similar to https://file-exchange.bihealth.org/<file-box-id>/. The files are served over an encrypted connection using WebDAV (which uses HTTPS).
The following describes how to access the files in the box from different platforms.
"},{"location":"how-to/service/file-exchange/#from-linux","title":"From Linux","text":"We describe how to access the files on the command line using the lftp program. The program is preinstalled on the BIH (and the MDC cluster) and you should be able to just install it with yum install lftp on CentOS/Red Hat or apt-get install lftp on Ubuntu/Debian.
When using lftp, you have to add some configuration first:
# cat >>~/.lftprc <<\"EOF\"\nset ssl:verify-certificate no\nset ftp:ssl-force yes\nEOF\nIn case that you want to access the files using a graphical user interface, search Google for \"WebDAV\" and your operating system or desktop environment. File browsers such as Nautilus and Thunar have built-in WebDAV support.
"},{"location":"how-to/service/file-exchange/#connecting","title":"Connecting","text":"First, log into the machine that has lftp installed. The login nodes of the BIH cluster do not have it installed but all compute and file transfer nodes have it. Go to the data download location.
host:~$ mkdir -p ~/scratch/download_dir\nhost:~$ cd ~/scratch/download_dir\nNext, start lftp. You can open the connection using open -u <user>@<DOMAIN> https://file-exchange.bihealth.org/<file-box-id>/ (NB: there is a trailing slash) where
<user> is your user name, e.g., holtgrem,<domain> is either MDC-BERLIN or CHARITE, and<file-box-id> the file box ID from the URL provided to you.When prompted, use your normal Charite/MDC password to login.
host:download_dir$ lftp\nlftp :~> open -u holtgrem@CHARITE https://file-exchange.bihealth.org/c62910b3-c1ba-49a5-81a6-a68f1f15aef6\nPassword:\ncd ok, cwd=/c62910b3-c1ba-49a5-81a6-a68f1f15aef6\nlftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6>\nYou can find a full reference of lftp on the lftp man page. You could also use help COMMAND on the lftp prompt. For example, to look at the files of the server for a bit...
lftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6> ls\ndrwxr-xr-x -- /\ndrwxr-xr-x -- dir\n-rw-r--r-- -- file1\nlftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6> find\n./\n./dir/\n./dir/file2\n./file1\nTo download all data use mirror, e.g. with -P 4 to use four download threads.
lftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6> mirror .\nTotal: 2 directories, 3 files, 0 symlinks\nNew: 3 files, 0 symlinks\nlftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6> exit\nhost:download_dir$ tree\n.\n\u251c\u2500\u2500 dir\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 file2\n\u251c\u2500\u2500 file1\n\u2514\u2500\u2500 file.txt\n\n1 directory, 3 files\nIgnoring gnutls_record_recv errors.
A common error to see is mirror: Fatal error: gnutls_record_recv: The TLS connection was non-properly terminated.. You can just ignore this.
To upload data, you can use mirror -R . which is essentially the \"reverse\" of the mirror command.
lftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6> mirror -R\nmirror: Fatal error: gnutls_record_recv: The TLS connection was non-properly terminated.\nmirror: Fatal error: gnutls_record_recv: The TLS connection was non-properly terminated.\nmirror: Fatal error: gnutls_record_recv: The TLS connection was non-properly terminated.\nTotal: 2 directories, 3 files, 0 symlinks\nModified: 3 files, 0 symlinks\n4 errors detected\nWe recommend to use WinSCP for file transfer.
After starting WinSCP, you will see a window titled Login. Just paste the URL (e.g., https://file-exchange.bihealth.org/c62910b3-c1ba-49a5-81a6-a68f1f15aef6/) of the file box into the Host name entry field. In this case, the fields File protocol etc. will be filled automatically. Next, enter your user name as user@CHARITE or user@MDC-BERLIN (the capitalization of the part behind the @ is important). The window should now look similar to the one below.
Proxy Configuration on Charite Network
If you are on the Charite network then you have to configure the proxy. Otherwise, you have to skip this step.
Click Advanced and a window titled Advanced Site Settings will pop up. Here, select Connection / Proxy in the left side. Select HTTP for the Proxy type. Then, enter proxy.charite.de as the Proxy host name and set the Port number to 8080. The window should nwo look as below. Then, click OK to apply the proxy settings.
Finally, click Login. You can now transfer files between the file exchange server and your local computer using drag and drop between WinSCP and your local Windows File Explorer. Alternatively, you can use the two-panel view of WinSCP to transfer files as described here.
For Mac, we you can also use lftp as described above in From Linux. You can find install instructions here online.
Proxy Configuration on Charite Network
If you are on the Charite network then you must have configured the proxy appropriately. Otherwise, you have to skip this step.
You can find them in your System Preference in the Network section, in the Advanced tab of your network (e.g., WiFi).
If you want to use a graphical interface then we recommend the usage of Cyberduck. After starting the program, click Open Connection on the top left, then select WebDAV (HTTPS) and fill out the form as in the following way. Paste the file box URL into the server field and use your login name (user@CHARITE or user@MDC-BERLIN) with your usual password.
If you need to perform access through a graphical user interface on your Mac, please contact hpc-helpdesk@bihealth.org for support.
"},{"location":"how-to/service/file-exchange/#security","title":"Security","text":"The file exchange server has the fail2ban software installed and configured (Charite, MDC, and BIH IPs are excluded from this).
If you are entering your user/password incorrectly for more than 5 times in 10 minutes then your machine will be banned for one hour. This means someone else that has the same IP address from the side of the file exchange server can get you blocked. This can happen if you are in the same home or university network with NAT or if you are behind a proxy. In this case you get a \"connection refused\" error. In this case, try again in one hour.
"},{"location":"how-to/software/apptainer/","title":"Using Apptainer (with Docker Images)","text":"Note
Singularity is now Apptainer! While Apptainer provides an singularity alias for backwards compatibility, it is recommanded to adapt all workflows to use the new binary apptainer.
Apptainer (https://apptainer.org/) is a popular alternative to docker, because it does not require to run as a privileged user. Apptainer can run Docker images out-of-the-box by converting them to the apptainer image format. The following guide gives a quick dive into using docker images with apptainer.
Build on your workstation, run on the HPC
Building images using Apptainer requires root privileges. We cannot give you these permissions on the BIH HPC. Thus, you will have to build the images on your local workstation (or anywhere where you have root access). You can then run the built images on the BIH HPC.
This is also true for the --writeable flag. Apparently it needs root permissions which you don't have on the cluster.
Link ~/.apptainer to ~/work/.apptainer
Because you only have a quota of 1 GB in your home directory, you should symlink ~/.apptainer to ~/work/.apptainer.
host:~$ mkdir -p ~/work/.apptainer && ln -sr ~/work/.apptainer ~/.apptainer\nIn case you already have a apptainer directory:
host:~$ mv ~/.apptainer ~/work/.apptainer && ln -sr ~/work/.apptainer ~/.apptainer\nRun a bash in a docker image:
host:~$ apptainer shell docker://godlovedc/lolcow\nRun a command in a docker image:
host:~$ apptainer exec docker://godlovedc/lolcow echo \"hello, hello!\"\nRun a bash in a docker image, enable access to the cuda driver (--nv) and mount a path (--bind or -B):
host:~$ apptainer shell --nv --bind /path_on_host/:/path_inside_container/ docker://godlovedc/lolcow\nCaveats
Notes
APPTAINERENV_: host:~$ APPTAINERENV_HELLO=123 apptainer shell docker://godlovedc/lolcow echo $HELLO\napptainer shell or apptainer exec uses as cwd the host callers cwd not the one set in the Dockerfile. One can change this by setting --pwd.The easiest variant to run a docker image available via a docker hub is by specifying its url. This causes apptainer to download the image and convert it to a apptainer image:
host:~$ apptainer run docker://godlovedc/lolcow\nor to open a shell inside the image
host:~$ apptainer shell docker://godlovedc/lolcow\nFurthermore, similar to docker, one can pull (and convert) remote image with the following call:
host:~$ apptainer pull docker://godlovedc/lolcow\nIn case your registry requires authentication you can provide it via a prompt by adding the option --docker-login:
host:~$ apptainer pull --docker-login docker://ilumb/mylolcow\nor by setting the following environment variables:
host:~$ export APPTAINER_DOCKER_USERNAME=ilumb\nhost:~$ export APPTAINER_DOCKER_PASSWORD=<redacted>\nhost:~$ apptainer pull docker://ilumb/mylolcow\nMore details can be found in the Apptainer documentation.
"},{"location":"how-to/software/apptainer/#option-2-converting-docker-images","title":"Option 2: Converting Docker Images","text":"Another option is to convert your docker image into the Apptainer/Singularity image format. This can be easily done using the docker images provided by docker2singularity.
To convert the docker image docker_image_name to the apptainer image apptainer_image_name one can use the following command line. The output image will be located in output_directory_for_images.
host:~$ docker run -v /var/run/docker.sock:/var/run/docker.sock -v /output_directory_for_images/:/output --privileged -t --rm quay.io/singularity/docker2singularity --name apptainer_image_name docker_image_name\nThe resulting image can then directly be used as image:
host:~$ apptainer exec apptainer_image_name.sif bash\nHere are some tips for making Docker images compatible with Apptainer taken from docker2singulrity:
~/.bashrc, ~/.profile, etc.ENTRYPOINT instruction set pointing to the command line interface to your pipeline.CMD - rely only on ENTRYPOINT.ENTRYPOINT docker run -i -t --entrypoint /bin/bash bids/example.--read-only --tmpfs /run --tmpfs /tmp parameters (this emulates the read-only behavior of Apptainer).USER instruction set.from the official website: \"Cell Ranger is a set of analysis pipelines that process Chromium single-cell RNA-seq output to align reads, generate feature-barcode matrices and perform clustering and gene expression analysis\"
"},{"location":"how-to/software/cell-ranger/#installation","title":"installation","text":"requires registration before download from here
to unpack Cell Ranger, its dependencies and the cellranger script:
cd /data/cephfs-1/home/users/$USER/work\nmv /path/to/cellranger-3.0.2.tar.gz .\ntar -xzvf cellranger-3.0.2.tar.gz\nwill be provided in /data/cephfs-1/work/projects/cubit/current/static_data/app_support/cellranger
add a file slurm.template to /data/cephfs-1/home/users/$USER/work/cellranger-3.0.2/martian-cs/v3.2.0/jobmanagers/sge.template with the following contents:
#!/usr/bin/env bash\n#\n# Copyright (c) 2016 10x Genomics, Inc. All rights reserved.\n#\n# =============================================================================\n# Setup Instructions\n# =============================================================================\n#\n# 1. Add any other necessary Slurm arguments such as partition (-p) or account\n# (-A). If your system requires a walltime (-t), 24 hours (24:00:00) is\n# sufficient. We recommend you do not remove any arguments below or Martian\n# may not run properly.\n#\n# 2. Change filename of slurm.template.example to slurm.template.\n#\n# =============================================================================\n# Template\n# =============================================================================\n#\n#SBATCH -J __MRO_JOB_NAME__\n#SBATCH --export=ALL\n#SBATCH --nodes=1 --ntasks-per-node=__MRO_THREADS__\n#SBATCH --signal=2\n#SBATCH --no-requeue\n#SBATCH --partition=medium\n#SBATCH --time=24:00:00\n### Alternatively: --ntasks=1 --cpus-per-task=__MRO_THREADS__\n### Consult with your cluster administrators to find the combination that\n### works best for single-node, multi-threaded applications on your system.\n#SBATCH --mem=__MRO_MEM_GB__G\n#SBATCH -o __MRO_STDOUT__\n#SBATCH -e __MRO_STDERR__\n\n__MRO_CMD__\nnote: on newer cellranger version, slurm.template needs to go to /data/cephfs-1/home/users/$USER/work/cellranger-XX/external/martian/jobmanagers/
if that hasn't been done yet, you can use cellranger mkfastq (details to be added)
count)","text":"create a script run_cellranger.sh with these contents (consult the documentation for help:
#!/bin/bash\n\n/data/cephfs-1/home/users/$USER/work/cellranger-3.0.2/cellranger count \\\n --id=sample_id \\\n --transcriptome=/data/cephfs-1/work/projects/cubit/current/static_data/app_support/cellranger/refdata-cellranger-${species}-3.0.0\\\n --fastqs=/path/to/fastqs \\\n --sample=sample_name \\\n --expect-cells=n_cells \\\n --jobmode=slurm \\\n --maxjobs=100 \\\n --jobinterval=1000\nand then submit the job via
sbatch --ntasks=1 --mem-per-cpu=4G --time=8:00:00 -p medium -o cellranger.log run_cellranger.sh\nadd a file sge.template to /data/cephfs-1/home/users/$USER/work/cellranger-3.0.2/martian-cs/v3.2.0/jobmanagers/sge.template with the following contents:
# =============================================================================\n# Template\n# =============================================================================\n#\n#$ -N __MRO_JOB_NAME__\n#$ -V\n#$ -pe smp __MRO_THREADS__\n#$ -cwd\n#$ -P medium\n#$ -o __MRO_STDOUT__\n#$ -e __MRO_STDERR__\n#$ -l h_vmem=__MRO_MEM_GB_PER_THREAD__G\n#$ -l h_rt=08:00:00\n\n#$ -m a\n#$ -M user@email.com\n\n__MRO_CMD__\nand submit the job via
qsub -cwd -V -pe smp 1 -l h_vmem=8G -l h_rt=24:00:00 -P medium -m a -j y run_cellranger.sh\nSSH Tunnels Considered Harmful
Please use our Open OnDemand Portal for running Jupyter notebooks!
The information below is still accurate. However, many users find it tricky to get SSH tunnels working correctly. A considerable number of parts is involved and you have to get each step 100% correct. Helpdesk cannot support you in problems with SSH tunnels that are caused by incorrect usage.
"},{"location":"how-to/software/jupyter/#what-is-jupyter","title":"What is Jupyter","text":"Project Jupyter is a networking protocol for interactive computing that allows the user to write and execute code for a high number of different programming languages. The most used client is Jupyter Notebook that can be encountered in various form all over the web. Its basic principle is a document consisting of different cells, each of which contains either code (executed in place) or documentation (written in markdown). This allows one to handily describe the processed workflow.
"},{"location":"how-to/software/jupyter/#setup-and-running-jupyter-on-the-cluster","title":"Setup and running Jupyter on the cluster","text":"Install Jupyter on the cluster (via conda, by creating a custom environment)
hpc-cpu-x:~$ conda create -n jupyter jupyter\nhpc-cpu-x:~$ conda activate jupyter\n(If you want to work in a language other than python, you can install more Jupyter language kernel, see the kernel list)
Now you can start the Jupyter server session (you may want to do this in a screen & srun --pty bash -i session as jupyter keeps running while you are doing computations)
hpc-cpu-x:~$ jupyter notebook --no-browser\nCheck the port number (usually 8888) in the on output and remember it for later:
[I 23:39:40.860 NotebookApp] The Jupyter Notebook is running at:\n[I 23:39:40.860 NotebookApp] http://localhost:8888/\nBy default, Jupyter will create an access token (a link stated in the output) to protect your notebook against unauthorized access which you have to save and enter in the accessing browser. You can change this to password base authorization via jupyter notebook password. If you are running multiple server on one or more nodes, one can separate them by changing the port number by adding --port=$PORT.
This is slightly trickier as we have to create a SSH connection/tunnel with potentially multiple hops in between. The easiest way is probably to configure your .ssh/config to automatically route your connection via the login node (and possibly MDC jail). This is described in our Advanced SSH config documentation
In short,add these lines to ~/.ssh/config (replace curly parts):
Host bihcluster\n user {USER_NAME}\n HostName hpc-login-2.cubi.bihealth.org\n\nHost hpc-cpu*\n user {USER_NAME}\n ProxyJump bihcluster\nFor MDC users outside the MDC network:
Host mdcjail\n HostName ssh1.mdc-berlin.de\n User {MDC_USER_NAME}\n\nHost bihcluster\n user {USER_NAME}\n HostName hpc-login-2.cubi.bihealth.org\n\nHost hpc-cpu*\n user {USER_NAME}\n ProxyJump bihcluster\nCheck that this config is working by connecting like this: ssh hpc-cpu-1. Please note that you cannot use any resources on this node without a valid Slurm session.
Now you setup a tunnel for your running Jupyter session:
workstation:~$ ssh -N -f -L 127.0.0.1:8888:localhost:{PORT} hpc-cpu-x\n8888. The cluster node srun has sent you to determines the last argument. You should now be able to connect to your Jupyter server by typing localhost:8888 in your webbrowser (see the note about token and password above).
It can and will happen that will lose connection, either due to network problems or due to shut-down of your computer. This is not a problem at all and you will not lose data, just reconnect to your session. If your notebooks are also losing connection (you will see a colorful remark in the top right corner), reconnect and click the colorful button. If this does not work, your work is still not lost as all cells that have been executed are automatically saved anyways. Copy all unexecuted cells (those are only saved periodically) and reload the browser page (after reconnecting) with F5. (you can also open a copy of the notebook in another tab, just be aware that there may be synchronisation problems)
There are two independent steps in ending a session:
Canceling the SSH tunnel
hpc-cpu-x:~$ ps aux | grep \"$PORT\"\nThis will give you something like this:
user 54 0.0 0.0 43104 784 ? Ss 15:06 0:00 ssh -N -f -L 127.0.0.1:8888:localhost:8888 hpc-cpu-x\nuser 58 0.0 0.0 41116 1024 tty1 S 15:42 0:00 grep --color=auto 8888\nfrom which you need the process ID (here 54)
hpc-cpu-x:~$ kill -9 $PID\nShutdown the Jupyter server
Open the Jupyter session, cancel the process with {Ctrl} + {C} and confirm {y}. Make sure you saved your notebooks beforehand (though auto-save catches most things).
"},{"location":"how-to/software/jupyter/#advanced","title":"Advanced","text":"If anyone has figured out, the following might also be interesting (please add):
Because the GPU nodes med030[1-4] has four GPU units we can train a model by using multiple GPUs in parallel. This How-To gives an example with Keras 2.2.4 together and tensorflow. Finally soem hints how you can submit a job on the cluster.
Hint
With tensorflow > 2.0 and newer keras version the multi_gpu_model is deprecated and you have to use the MirroredStrategy.
we need to import the multi_gpu_model model from keras.utils and have to pass our actual model (maybe sequential Keras model) into it. In general Keras automatically configures the number of available nodes (gpus=None). This seems not to work on our system. So we have to specify the numer of GPUs, e.g. two with gpus=2. We put this in a try catch environment that it will also work on CPUs.
from keras.utils import multi_gpu_model\n\ntry: \n model = multi_gpu_model(model, gpus=2) \nexcept:\n pass\nThat's it!
Please read here on how to submit jobs to the GPU nodes.
"},{"location":"how-to/software/keras/#conda-environment","title":"Conda environment","text":"All this was tested with the following conda environment:
name: cuda channels: \n- conda-forge\n- bioconda\n- defaults\ndependencies:\n- keras=2.2.4\n- python=3.6.7\n- tensorboard=1.12.0\n- tensorflow=1.12.0\n- tensorflow-base=1.12.0\n- tensorflow-gpu=1.12.0\nNote
This information is outdated and will soon be removed.
GNU Octave as Matlab alternative
Note that GNU Octave is an Open Source alternative to Matlab. While both packages are not 100% compatible, Octave is an alternative that does not require any license management. Further, you can easily install it yourself using Conda.
Want to use the Matlab GUI?
Make sure you understand X forwarding as outline in this FAQ entry.
You can also use Open OnDemand Portal to run Matlab.
"},{"location":"how-to/software/matlab/#pre-requisites","title":"Pre-requisites","text":"You have to register with hpc-helpdesk@bih-charite.de for requesting access to the Matlab licenses. Afterwards, you can connect to the High-Memory using the license_matlab_r2016b resource (see below).
BIH has a license of Matlab R2016b for 16 seats and various licensed packages (see below). To display the available licenses:
hpc-login-1:~$ scontrol show lic\nLicenseName=matlab_r2016b\n Total=16 Used=0 Free=16 Remote=no\nMatlab is installed on all of the compute nodes:
# The following is VITAL so the scheduler allocates a license to your session.\nhpc-login-1:~$ srun -L matlab_r2016b:1 --pty bash -i\nmed0127:~$ scontrol show lic\nLicenseName=matlab_r2016b\n Total=16 Used=1 Free=15 Remote=no\nmed0127:~$ module avail\n----------------- /usr/share/Modules/modulefiles -----------------\ndot module-info null\nmodule-git modules use.own\n\n----------------------- /opt/local/modules -----------------------\ncmake/3.11.0-0 llvm/6.0.0-0 openmpi/3.1.0-0\ngcc/7.2.0-0 matlab/r2016b-0\nmed0127:~$ module load matlab/r2016b-0\nStart matlab without GUI: matlab -nosplash -nodisplay -nojvm\n Start matlab with GUI (requires X forwarding (ssh -X)): matlab\nmed0127:~$ matlab -nosplash -nodisplay -nojvm\n < M A T L A B (R) >\n Copyright 1984-2016 The MathWorks, Inc.\n R2016b (9.1.0.441655) 64-bit (glnxa64)\n September 7, 2016\n\n\nFor online documentation, see http://www.mathworks.com/support\nFor product information, visit www.mathworks.com.\n\n\n Non-Degree Granting Education License -- for use at non-degree granting, nonprofit,\n educational organizations only. Not for government, commercial, or other organizational use.\n\n>> ver\n--------------------------------------------------------------------------------------------\nMATLAB Version: 9.1.0.441655 (R2016b)\nMATLAB License Number: 1108905\nOperating System: Linux 3.10.0-862.3.2.el7.x86_64 #1 SMP Mon May 21 23:36:36 UTC 2018 x86_64\nJava Version: Java is not enabled\n--------------------------------------------------------------------------------------------\nMATLAB Version 9.1 (R2016b)\nBioinformatics Toolbox Version 4.7 (R2016b)\nGlobal Optimization Toolbox Version 3.4.1 (R2016b)\nImage Processing Toolbox Version 9.5 (R2016b)\nOptimization Toolbox Version 7.5 (R2016b)\nParallel Computing Toolbox Version 6.9 (R2016b)\nPartial Differential Equation Toolbox Version 2.3 (R2016b)\nSignal Processing Toolbox Version 7.3 (R2016b)\nSimBiology Version 5.5 (R2016b)\nStatistics and Machine Learning Toolbox Version 11.0 (R2016b)\nWavelet Toolbox Version 4.17 (R2016b)\n>> exit\nFor starting the Matlab with GUI, make sure that your client is running a X11 server and you connect with X11 forwarding enabled (e.g., ssh -X hpc-login-1.cubi.bihealth.org from the Linux command line). Then, make sure to use srun -L matlab_r2016b:1 --pty --x11 bash -i for connecting to a node with X11 forwarding enabled.
client:~$ ssh -X hpc-login-1.cubi.bihealth.org\n[...]\nhpc-login-1:~ $ srun -L matlab_r2016b:1 --pty --x11 bash -i\n[...]\nmed0203:~$ module load matlab/r2016b-0\nStart matlab without GUI: matlab -nosplash -nodisplay -nojvm\n Start matlab with GUI (requires X forwarding (ssh -X)): matlab\nmed0203:~$ matlab\n[UI will start]\nFor forcing starting in text mode can be done (as said after module load): matlab -nosplash -nodisplay -nojvm.
Also see this FAQ entry.
"},{"location":"how-to/software/matlab/#see-available-matlab-licenses","title":"See Available Matlab Licenses","text":"You can use scontrol show lic to see the currently available MATLAB license. E.g., here I am running an interactive shell in which I have requested 1 of the 16 MATLAB licenses, so 15 more remain.
$ scontrol show lic\nLicenseName=matlab_r2016b\n Total=16 Used=1 Free=15 Remote=no\nGet a checkout of our MATLAB example. Then, look around at the contents of this repository.
hpc-login-1:~$ git clone https://github.com/bihealth/bih-cluster-matlab-example.git\nhpc-login-1:~$ cd bih-cluster-matlab-example\nhpc-login-1:~$ cat job_script.sh\n#!/bin/bash\n\n# Logging goes to directory sge_log\n#SBATCH -o slurm_log/%x-%J.log\n# Keep current environment variables\n#SBATCH --export=ALL\n# Name of the script\n#SBATCH --job-name MATLAB-example\n# Allocate 4GB of RAM per core\n#SBATCH --mem 4G\n# Maximal running time of 2 hours\n#SBATCH --time 02:00:00\n# Allocate one Matlab license\n#SBATCH -L matlab_r2016b:1\n\nmodule load matlab/r2016b-0\n\nmatlab -r example\n$ cat example.m\n% Example Hello World script for Matlab.\n\ndisp('Hello world!')\ndisp('Thinking...')\n\npause(10)\n\ndisp(sprintf('The square root of 2 is = %f', sqrt(2)))\nexit\nFor submitting the script, you can do the following
hpc-login-1:~$ sbatch job_script.sh\nThis will submit a job with one Matlab license requested. If you were to submit 17 of these jobs, then at least one of them would have to wait until all licenses are free.
Matlab License Server
Note that there is a Matlab license server running on the server that will check whether 16 or less Matlab sessions are currently running. If a Matlab session is running but this is not made known to the scheduler via -L matlab_r2016b then this can lead to scripts crashing as not enough licenses are available. If this happens to you, double-check that you have specified the license requirements correctly and notify hpc-helpdesk@bih-charite.de in case of any problems. We will try to sort out the situation then.
This article describes how to build an run an OpenMPI program. We will build a simple C program that uses the OpenMPI message passing interface and run it in parallel. You should be able to go from here with other languages and more complex programs. We will use a simple Makefile for building the software.
"},{"location":"how-to/software/openmpi/#loading-openmpi-environment","title":"Loading OpenMPI Environment","text":"First, load the OpenMPI package.
hpc-login-1:~$ srun --pty bash -i\nmed0127:~$ module load openmpi/4.3.0-0\nThen, check that the installation works
med0127:~$ ompi_info | head\n Package: Open MPI root@med0127 Distribution\n Open MPI: 4.0.3\n Open MPI repo revision: v4.0.3\n Open MPI release date: Mar 03, 2020\n Open RTE: 4.0.3\n Open RTE repo revision: v4.0.3\n Open RTE release date: Mar 03, 2020\n OPAL: 4.0.3\n OPAL repo revision: v4.0.3\n OPAL release date: Mar 03, 2020\nNext, clone the OpenMPI example project from Gitlab.
med0127:~$ git clone git@github.com:bihealth/bih-cluster-openmpi-example.git\nmed0127:~$ cd bih-cluster-openmpi-example/src\nMakefile
.PHONY: default clean\n\n# configure compilers\nCC=mpicc\nCXX=mpicxx\n# configure flags\nCCFLAGS += $(shell mpicc --showme:compile)\nLDFLAGS += $(shell mpicc --showme:link)\n\ndefault: openmpi_example\n\nopenmpi_example: openmpi_example.o\n\nclean:\n rm -f openmpi_example.o openmpi_example\nopenmpi_example.c
#include <stdio.h>\n#include <mpi.h>\n\nint main(int argc, char** argv) {\n // Initialize the MPI environment\n MPI_Init(NULL, NULL);\n\n // Get the number of processes\n int world_size;\n MPI_Comm_size(MPI_COMM_WORLD, &world_size);\n\n // Get the rank of the process\n int world_rank;\n MPI_Comm_rank(MPI_COMM_WORLD, &world_rank);\n\n // Get the name of the processor\n char processor_name[MPI_MAX_PROCESSOR_NAME];\n int name_len;\n MPI_Get_processor_name(processor_name, &name_len);\n\n // Print off a hello world message\n printf(\"Hello world from processor %s, rank %d\"\n \" out of %d processors\\n\",\n processor_name, world_rank, world_size);\n\n // Finalize the MPI environment.\n MPI_Finalize();\n\n return 0;\n}\nrun_mpi.sh
#!/bin/bash\n\n# Example job script for (single-threaded) MPI programs.\n\n# Generic arguments\n\n# Job name\n#SBATCH --job-name openmpi_example\n# Maximal running time of 10 min\n#SBATCH --time 00:10:00\n# Allocate 1GB of memory per node\n#SBATCH --mem 1G\n# Write logs to directory \"slurm_log\"\n#SBATCH -o slurm_log/slurm-%x-%J.log\n\n# MPI-specific parameters\n\n# Run 64 tasks (threads/on virtual cores)\n#SBATCH --nodes 64\n\n# Make sure to source the profile.d file (not available on head nodes).\n/etc/profile.d/modules.sh\n\n# Load the OpenMPI environment module to get the runtime environment.\nmodule load openmpi/3.1.0-0\n\n# Launch the program.\nmpirun -np 64 ./openmpi_example\nThe next step is building the software
med0127:~$ make\nmpicc -c -o openmpi_example.o openmpi_example.c\nmpicc -pthread -Wl,-rpath -Wl,/opt/local/openmpi-4.0.3-0/lib -Wl,--enable-new-dtags -L/opt/local/openmpi-4.0.3-0/lib -lmpi openmpi_example.o -o openmpi_example\nmed0127:~$ ls -lh\ntotal 259K\n-rw-rw---- 1 holtgrem_c hpc-ag-cubi 287 Apr 7 23:29 Makefile\n-rwxrwx--- 1 holtgrem_c hpc-ag-cubi 8.5K Apr 8 00:15 openmpi_example\n-rw-rw---- 1 holtgrem_c hpc-ag-cubi 760 Apr 7 23:29 openmpi_example.c\n-rw-rw---- 1 holtgrem_c hpc-ag-cubi 2.1K Apr 8 00:15 openmpi_example.o\n-rwxrwx--- 1 holtgrem_c hpc-ag-cubi 1.3K Apr 7 23:29 run_hybrid.sh\n-rwxrwx--- 1 holtgrem_c hpc-ag-cubi 663 Apr 7 23:35 run_mpi.sh\ndrwxrwx--- 2 holtgrem_c hpc-ag-cubi 4.0K Apr 7 23:29 sge_log\nThe software will run outside of the MPI environment -- but in a single process only, of course.
med0127:~$ ./openmpi_example\nHello world from processor med0127, rank 0 out of 1 processors\nAll of the arguments are already in the run_mpi.sh script.
med01247:~# sbatch run_mpi.sh\nExplanation of the OpenMPI-specific arguments
--ntasks 64: run 64 processes in the MPI environment.Let's look at the slurm log file, e.g., in slurm_log/slurm-openmpi_example-3181.log.
med0124:~$ cat slurm_log/slurm-openmpi_example-*.log\nHello world from processor med0133, rank 6 out of 64 processors\nHello world from processor med0133, rank 25 out of 64 processors\nHello world from processor med0133, rank 1 out of 64 processors\nHello world from processor med0133, rank 2 out of 64 processors\nHello world from processor med0133, rank 3 out of 64 processors\nHello world from processor med0133, rank 7 out of 64 processors\nHello world from processor med0133, rank 9 out of 64 processors\nHello world from processor med0133, rank 12 out of 64 processors\nHello world from processor med0133, rank 13 out of 64 processors\nHello world from processor med0133, rank 15 out of 64 processors\nHello world from processor med0133, rank 16 out of 64 processors\nHello world from processor med0133, rank 17 out of 64 processors\nHello world from processor med0133, rank 18 out of 64 processors\nHello world from processor med0133, rank 23 out of 64 processors\nHello world from processor med0133, rank 24 out of 64 processors\nHello world from processor med0133, rank 26 out of 64 processors\nHello world from processor med0133, rank 27 out of 64 processors\nHello world from processor med0133, rank 31 out of 64 processors\nHello world from processor med0133, rank 0 out of 64 processors\nHello world from processor med0133, rank 4 out of 64 processors\nHello world from processor med0133, rank 5 out of 64 processors\nHello world from processor med0133, rank 8 out of 64 processors\nHello world from processor med0133, rank 10 out of 64 processors\nHello world from processor med0133, rank 11 out of 64 processors\nHello world from processor med0133, rank 14 out of 64 processors\nHello world from processor med0133, rank 19 out of 64 processors\nHello world from processor med0133, rank 20 out of 64 processors\nHello world from processor med0133, rank 21 out of 64 processors\nHello world from processor med0133, rank 22 out of 64 processors\nHello world from processor med0133, rank 28 out of 64 processors\nHello world from processor med0133, rank 29 out of 64 processors\nHello world from processor med0133, rank 30 out of 64 processors\nHello world from processor med0134, rank 32 out of 64 processors\nHello world from processor med0134, rank 33 out of 64 processors\nHello world from processor med0134, rank 34 out of 64 processors\nHello world from processor med0134, rank 38 out of 64 processors\nHello world from processor med0134, rank 39 out of 64 processors\nHello world from processor med0134, rank 42 out of 64 processors\nHello world from processor med0134, rank 44 out of 64 processors\nHello world from processor med0134, rank 45 out of 64 processors\nHello world from processor med0134, rank 46 out of 64 processors\nHello world from processor med0134, rank 53 out of 64 processors\nHello world from processor med0134, rank 54 out of 64 processors\nHello world from processor med0134, rank 55 out of 64 processors\nHello world from processor med0134, rank 60 out of 64 processors\nHello world from processor med0134, rank 62 out of 64 processors\nHello world from processor med0134, rank 35 out of 64 processors\nHello world from processor med0134, rank 36 out of 64 processors\nHello world from processor med0134, rank 37 out of 64 processors\nHello world from processor med0134, rank 40 out of 64 processors\nHello world from processor med0134, rank 41 out of 64 processors\nHello world from processor med0134, rank 43 out of 64 processors\nHello world from processor med0134, rank 47 out of 64 processors\nHello world from processor med0134, rank 48 out of 64 processors\nHello world from processor med0134, rank 49 out of 64 processors\nHello world from processor med0134, rank 50 out of 64 processors\nHello world from processor med0134, rank 51 out of 64 processors\nHello world from processor med0134, rank 52 out of 64 processors\nHello world from processor med0134, rank 56 out of 64 processors\nHello world from processor med0134, rank 57 out of 64 processors\nHello world from processor med0134, rank 59 out of 64 processors\nHello world from processor med0134, rank 61 out of 64 processors\nHello world from processor med0134, rank 63 out of 64 processors\nHello world from processor med0134, rank 58 out of 64 processors\nIn some cases, you want to mix multithreading (e.g., via OpenMP) with MPI to run one process with multiple threads that then can communicate via shared memory. Note that OpenMPI will let processes on the same node communicate via shared memory anyway, so this might not be necessary in all cases.
The file run_hybrid.sh shows how to run an MPI job with 8 threads each.
Note well that memory is allocated on a per-slot (thus per-thread) base!
run_hybrid.sh
#!/bin/bash\n\n# Example job script for multi-threaded MPI programs, sometimes\n# called \"hybrid\" MPI computing.\n\n# Generic arguments\n\n# Job name\n#SBATCH --job-name openmpi_example\n# Maximal running time of 10 min\n#SBATCH --time 00:10:00\n# Allocate 1GB of memory per node\n#SBATCH --mem 1G\n# Write logs to directory \"slurm_log\"\n#SBATCH -o slurm_log/slurm-%x-%J.log\n\n# MPI-specific parameters\n\n# Run 8 tasks (threads/on virtual cores)\n#SBATCH --ntasks 8\n# Allocate 4 CPUs per task (cores/threads)\n#SBATCH --cpus-per-task 4\n\n# Make sure to source the profile.d file (not available on head nodes).\nsource /etc/profile.d/modules.sh\n\n# Load the OpenMPI environment module to get the runtime environment.\nmodule load openmpi/4.0.3-0\n\n# Launch the program.\nmpirun -n 8 ./openmpi_example\nWe changed the following
Let's look at the log output:
# cat slurm_log/slurm-openmpi_example-3193.log\nHello world from processor med0133, rank 1 out of 8 processors\nHello world from processor med0133, rank 3 out of 8 processors\nHello world from processor med0133, rank 2 out of 8 processors\nHello world from processor med0133, rank 6 out of 8 processors\nHello world from processor med0133, rank 0 out of 8 processors\nHello world from processor med0133, rank 4 out of 8 processors\nHello world from processor med0133, rank 5 out of 8 processors\nHello world from processor med0133, rank 7 out of 8 processors\nEach process can now launch 4 threads (e.g., by defining export OMP_NUM_THREADS=4 before the program call).
This page gives an end-to-end example how to build and install Gromacs as an example for managing complex scientific software installs in user land. You don't have to learn or understand the specifics of Gromacs. We use it as an example as there are some actual users on the BIH cluster. However, installing it is out of scope of BIH HPC administration.
Gromacs is a good example as it is a sufficiently complex piece of software. Quite some configuration is done on the command line and there is no current software package of it in the common RPM repositories. However, it is quite well-documented and easy to install for scientific software so there is a lot to be learned.
"},{"location":"how-to/software/scientific-software/#related-documents","title":"Related Documents","text":"We will perform the following step:
Makefiles)Many scientific software packages will have more dependencies. If the dependencies are available as CentOS Core or EPEL packages (such as zlib), HPC IT administration can install them. However, otherwise you will have to install them on their own.
Warning
Do not perform the compilation on the login nodes but go to a compute node instead.
"},{"location":"how-to/software/scientific-software/#downloading-and-extracting-software","title":"Downloading and Extracting Software","text":"This is best done in your scratch directory as we don't have to keep these files around for long. Note that the files in your scratch directory will automatically be removed after 2 weeks. You can also use your work directory here.
hpc-login-1:~$ srun --pty bash -i\nmed0127:~$ mkdir $HOME/scratch/gromacs-install\nmed0127:~$ cd $HOME/scratch/gromacs-install\nmed0127:~$ wget http://ftp.gromacs.org/pub/gromacs/gromacs-2018.3.tar.gz\nmed0127:~$ tar xf gromacs-2018.3.tar.gz\nmed0127:~$ ls gromacs-2018.3\nadmin cmake COPYING CTestConfig.cmake INSTALL scripts src\nAUTHORS CMakeLists.txt CPackInit.cmake docs README share tests\nSo far so good!
"},{"location":"how-to/software/scientific-software/#perform-the-configure-step","title":"Perform the Configure Step","text":"This is the most critical step. Most scientific C/C++ software has a build step and allows for, e.g., disabling and enabling features or setting installation paths. Here, you can configure the software depending on your needs and environment. Also, it is the easiest step to mess up.
Gromac's documentation is actually quite good but the author had problems to follow it to the letter. Gromacs recommends to create an MPI and a non-MPI build but the precise way did not work. This installation creates two flavours for Gromacs 2018.3, but in a different way than the Gromacs documentation proposes.
First, here is how to configure the non-MPI flavour Gromacs wants a modern compiler, so we load gcc. We will need to note down the precise version we used so later we can load it for running Gromacs with the appropriate libraries. We will install gromacs into $HOME/work/software, which is appropriate for user-installed software, but it could also go into a group or project directory. Note that we install the software into your work directory as software installations are quite large and might go above your home quota. Also, software installations are usually not precious enough to waste resources on snapshots and backups. Also that we force Gromacs to use AVX_256 for SIMD support (Intel sandy bridge architecture) to not get unsupported CPU instruction errors.
med0127:~$ module load gcc/7.2.0-0 cmake/3.11.0-0\nmed0127:~$ module list\nCurrently Loaded Modulefiles:\n 1) gcc/7.2.0-0 2) cmake/3.11.0-0\nmed0127:~$ mkdir gromacs-2018.3-build-nompi\nmed0127:~$ cd gromacs-2018.3-build-nompi\nmed0127:~$ cmake ../gromacs-2018.3 \\\n -DGMX_BUILD_OWN_FFTW=ON \\\n -DGMX_MPI=OFF \\\n -DGMX_SIMD=AVX_256 \\\n -DCMAKE_INSTALL_PREFIX=$HOME/work/software/gromacs/2018.3\nSecond, here is how to configure the MPI flavour. Note that we are also enabling the openmpi module. We will also need the precise version here so we can later load the correct libraries. Note that we install the software into the directory gromacs-mpi but switch off shared library building as recommended by the Gromacs documentation.
med0127:~$ module load openmpi/3.1.0-0\nmed0127:~$ module list\nCurrently Loaded Modulefiles:\n 1) gcc/7.2.0-0 2) cmake/3.11.0-0 3) openmpi/4.0.3-0\nmed0127:~$ mkdir gromacs-2018.3-build-mpi\nmed0127:~$ cd gromacs-2018.3-build-mpi\nmed0127:~$ cmake ../gromacs-2018.3 \\\n -DGMX_BUILD_OWN_FFTW=ON \\\n -DGMX_MPI=ON \\\n -DGMX_SIMD=AVX_256 \\\n -DCMAKE_INSTALL_PREFIX=$HOME/work/software/gromacs-mpi/2018.3 \\\n -DCMAKE_C_COMPILER=$(which mpicc) \\\n -DCMAKE_CXX_COMPILER=$(which mpicxx) \\\n -DBUILD_SHARED_LIBS=off\nThis is simple, using -j 32 allows us to build with 32 threads. If something goes wrong: meh, the \"joys\" of compilling C software.
Getting Support for Building Software
BIH HPC IT cannot provide support for compiling scientific software. Please contact the appropriate mailing lists or forums for your scientific software. You should only contact the BIH HPC IT helpdesk only if you are sure that the problem is with the BIH HPC cluster. You should try to resolve the issue on your own and with the developers of the software that you are trying to build/use.
For the no-MPI version:
med0127:~$ cd ../cd gromacs-2018.3-build-nompi\nmed0127:~$ make -j 32\n[...]\nmed0127:~$ make install\nFor the MPI version:
med0127:~$ cd ../cd gromacs-2018.3-build-mpi\nmed0127:~$ make -j 32\n[...]\nmed0127:~$ make install\nFor Gromacs 2018.3, the following is appropriate. You should be able to use this as a template for your environment module files:
med0127:~$ mkdir -p $HOME/local/modules/gromacs\nmed0127:~$ cat >$HOME/local/modules/gromacs/2018.3 <<\"EOF\"\n#%Module\nproc ModulesHelp { } {\n puts stderr {\n Gromacs molecular simulation toolkit (non-MPI version)\n\n - http://www.gromacs.org\n }\n}\n\nmodule-whatis {Gromacs molecular simulation toolkit (non-MPI)}\n\nset root /data/cephfs-1/home/users/YOURUSER/work/software/gromacs-mpi/2018.3\n\nprereq gcc/7.2.0-0\n\nconflict gromacs\nconflict gromacs-mpi\n\nprepend-path LD_LIBRARY_PATH $root/lib64\nprepend-path LIBRARY_PATH $root/lib64\nprepend-path MANPATH $root/share/man\nprepend-path PATH $root/bin\nsetenv GMXRC $root/bin/GMXRC\nEOF\nmed0127:~$ mkdir -p $HOME/local/modules/gromacs-mpi\nmed0127:~$ cat >$HOME/local/modules/gromacs-mpi/2018.3 <<\"EOF\"\n#%Module\nproc ModulesHelp { } {\n puts stderr {\n Gromacs molecular simulation toolkit (MPI version)\n\n - http://www.gromacs.org\n }\n}\n\nmodule-whatis {Gromacs molecular simulation toolkit (MPI)}\n\nset root /data/cephfs-1/home/users/YOURUSER/work/software/gromacs-mpi/2018.3\n\nprereq openmpi/4.0.3-0\nprereq gcc/7.2.0-0\n\nconflict gromacs\nconflict gromacs-mpi\n\nprepend-path LD_LIBRARY_PATH $root/lib64\nprepend-path LIBRARY_PATH $root/lib64\nprepend-path MANPATH $root/share/man\nprepend-path PATH $root/bin\nsetenv GMXRC $root/bin/GMXRC\nEOF\nWith the next command, make your local modules files path known to the environemtn modules system.
med0127:~$ module use $HOME/local/modules\nYou can verify the result:
med0127:~$ module avail\n\n------------------ /data/cephfs-1/home/users/YOURUSER/local/modules ------------------\ngromacs/2018.3 gromacs-mpi/2018.3\n\n-------------------- /usr/share/Modules/modulefiles --------------------\ndot module-info null\nmodule-git modules use.own\n\n-------------------------- /opt/local/modules --------------------------\ncmake/3.11.0-0 llvm/6.0.0-0 openmpi/3.1.0-0\ngcc/7.2.0-0 matlab/r2016b-0 openmpi/4.0.3-0\nmodule use","text":"You can add this to your ~/.bashrc file to always execute the module use after login. Note that module is not available on the login or transfer nodes, the following should work fine:
med0127:~$ cat >>~/.bashrc <<\"EOF\"\ncase \"${HOSTNAME}\" in\n login-*|transfer-*)\n ;;\n *)\n module use $HOME/local/modules\n ;;\nesac\nEOF\nNote that the paths chosen above are sensible but arbitrary. You can install any software anywhere you have permission to -- somewhere in your user and group home, maybe a project home makes most sense on the BIH HPC, no root permissions required. You can also place the module files anywhere, as long as the module use line is appropriate.
As a best practice, you could use the following location:
$HOME/work/software as a root to install software to$HOME/work/software/$PKG/$VERSION for installing a given software package in a given version$HOME/work/software/modules as the root for modules to install$HOME/work/software/$PKG/$VERSION for the module file to load the software in a given version$HOME/work/software/modules.sh as a Bash script to contain the line module use $HOME/work/software/moduleschmod ug=rX,o= $GROUP/work/software, the upper case X is essential to only set +x on directories and not files):$GROUP/work/software as a root to install software to$GROUP/work/software/$PKG/$VERSION for installing a given software package in a given version$GROUP/work/software/modules as the root for modules to install$GROUP/work/software/$PKG/$VERSION for the module file to load the software in a given version$GROUP/work/software/modules.sh as a Bash script to contain the case Bash snippet from above but with module use $GROUP/work/software/modulesEvery time you want to use Gromacs, you can now do
med0127:~$ module load gcc/7.2.0-0 gromacs/2018.3\nor, if you want to have the MPI version:
med0127:~$ module load gcc/7.2.0-0 openmpi/4.0.3-0 gromacs-mpi/2018.3\nSomething along the lines of the following job script should be appropriate. See How-To: Build Run OpenMPI Programs for more information.
#!/bin/bash\n\n# Example job script for (single-threaded) MPI programs.\n\n# Generic arguments\n\n# Job name\n#SBATCH --job-name gromacs\n# Maximal running time of 10 min\n#SBATCH --time 00:10:00\n# Allocate 1GB of memory per CPU\n#SBATCH --mem 1G\n# Write logs to directory \"slurm_log/<name>-<job id>.log\" (dir must exist)\n#SBATCH --output slurm_log/%x-%J.log\n\n# MPI-specific parameters\n\n# Launch on 8 nodes (== 8 tasks)\n#SBATCH --ntasks 8\n# Allocate 4 CPUs per task (== per node)\n#SBATCH --cpus-per-task 4\n\n# Load the OpenMPI and GCC environment module to get the runtime environment.\nmodule load gcc/4.7.0-0\nmodule load openmpi/4.0.3-0\n\n# Make custom environment modules known. Alternative, you can \"module use\"\n# them in the session you use for submitting the job.\nmodule use $HOME/local/modules\nmodule load gromacs-mpi/2018.3\n\n# Launch the program on 8 nodes and tell Gromacs to use 4 threads for each\n# invocation.\nexport OMP_NUM_THREADS=4\nmpirun -n 8 gmx_mpi mdrun -deffnm npt_1000\nmed0127:~$ mkdir slurm_log\nmed0127:~$ sbatch job_script.sh\nSubmitted batch job 3229\nTensorFlow is a package for deep learning with optional support for GPUs. You can find the original TensorFlow installation instructions here.
This article describes how to set up TensorFlow with GPU support using Conda. This how-to assumes that you have just connected to a GPU node via srun --mem=10g --partition=gpu --gres=gpu:tesla:1 --pty bash -i (for Tesla V100 GPUs, for A400 GPUs use --gres=gpu:a40:1). Note that you will need to allocate \"enough\" memory, otherwise your python session will be Killed because of too little memory. You should read the How-To: Connect to GPU Nodes tutorial on an explanation of how to do this.
This tutorial assumes, that conda has been set up as described in [Software Management]((../../best-practice/software-installation-with-conda.md).
"},{"location":"how-to/software/tensorflow/#create-conda-environment","title":"Create conda environment","text":"We recommend that you install mamba first with conda install -y mamba and use this C++ reimplementation of the conda command as follows.
$ conda create -y -n python-tf tensorflow-gpu\n$ conda activate python-tf\nLet us verify that we have Python and TensorFlow installed. You might get different versions you could pin the version on installing with `conda create -y -n python-tf python==3.9.10 tensorflow-gpu==2.6.2
$ python --version\nPython 3.9.10\n$ python -c 'import tensorflow; print(tensorflow.__version__)'\n2.6.2\nWe thus end up with an installation of Python 3.9.10 with tensorflow 2.6.2.
"},{"location":"how-to/software/tensorflow/#run-tensorflow-example","title":"Run TensorFlow Example","text":"Let us now see whether TensorFlow has recognized our GPU correctly.
$ python\n>>> import tensorflow as tf\n>>> print(\"TensorFlow version:\", tf.__version__)\nTensorFlow version: 2.6.2\n>>> print(tf.config.list_physical_devices())\n[PhysicalDevice(name='/physical_device:CPU:0', device_type='CPU'), PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]\nYay, we can proceed to run the Quickstart Tutorial.
>>> mnist = tf.keras.datasets.mnist\n>>> (x_train, y_train), (x_test, y_test) = mnist.load_data()\n>>> x_train, x_test = x_train / 255.0, x_test / 255.0\n>>> model = tf.keras.models.Sequential([\n... tf.keras.layers.Flatten(input_shape=(28, 28)),\n... tf.keras.layers.Dense(128, activation='relu'),\n... tf.keras.layers.Dropout(0.2),\n... tf.keras.layers.Dense(10)\n... ])\n>>> predictions = model(x_train[:1]).numpy()\n>>> predictions\narray([[-0.50569224, 0.26386747, 0.43226188, 0.61226094, 0.09630793,\n 0.34400576, 0.9819117 , -0.3693726 , 0.5221357 , 0.3323232 ]],\n dtype=float32)\n>>> tf.nn.softmax(predictions).numpy()\narray([[0.04234391, 0.09141268, 0.10817807, 0.12951255, 0.07731011,\n 0.09903987, 0.18743432, 0.04852816, 0.11835073, 0.09788957]],\n dtype=float32)\n>>> loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)\n>>> loss_fn(y_train[:1], predictions).numpy()\n2.3122327\n>>> model.compile(optimizer='adam',\n... loss=loss_fn,\n... metrics=['accuracy'])\n>>> model.fit(x_train, y_train, epochs=5)\n2022-03-09 17:53:47.237997: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:185] None of the MLIR Optimization Passes are enabled (registered 2)\nEpoch 1/5\n1875/1875 [==============================] - 3s 1ms/step - loss: 0.2918 - accuracy: 0.9151\nEpoch 2/5\n1875/1875 [==============================] - 3s 1ms/step - loss: 0.1444 - accuracy: 0.9561\nEpoch 3/5\n1875/1875 [==============================] - 3s 1ms/step - loss: 0.1082 - accuracy: 0.9674\nEpoch 4/5\n1875/1875 [==============================] - 3s 1ms/step - loss: 0.0898 - accuracy: 0.9720\nEpoch 5/5\n1875/1875 [==============================] - 3s 1ms/step - loss: 0.0773 - accuracy: 0.9756\n<keras.callbacks.History object at 0x154e81360190>\n>>> model.evaluate(x_test, y_test, verbose=2)\n313/313 - 0s - loss: 0.0713 - accuracy: 0.9785\n[0.0713074803352356, 0.9785000085830688]\n>>> probability_model = tf.keras.Sequential([\n... model,\n... tf.keras.layers.Softmax()\n... ])\n>>> probability_model(x_test[:5])\n<tf.Tensor: shape=(5, 10), dtype=float32, numpy=\narray([[1.2339272e-06, 6.5599060e-10, 1.0560590e-06, 5.9356184e-06,\n 5.3691075e-12, 1.4447859e-07, 5.4218874e-13, 9.9996936e-01,\n 1.0347234e-07, 2.2147648e-05],\n [2.9887938e-06, 6.8461006e-05, 9.9991941e-01, 7.2003731e-06,\n 2.9751782e-13, 8.2818183e-08, 1.4307782e-06, 2.3203837e-13,\n 4.7433215e-07, 2.9504194e-14],\n [1.8058477e-06, 9.9928612e-01, 7.8716243e-05, 3.9140195e-06,\n 3.0842333e-05, 9.4537208e-06, 2.2774333e-05, 4.5549971e-04,\n 1.1015874e-04, 6.9138093e-07],\n [9.9978787e-01, 3.0206781e-08, 2.8528208e-05, 8.5581682e-08,\n 1.3851340e-07, 2.3634559e-06, 1.8480707e-05, 1.0153375e-04,\n 1.1583331e-07, 6.0887167e-05],\n [6.4914235e-07, 2.5808356e-08, 1.8225538e-06, 2.3215563e-09,\n 9.9588013e-01, 4.6049720e-08, 3.8903639e-07, 2.9772724e-05,\n 4.3141077e-07, 4.0867776e-03]], dtype=float32)>\n>>> exit()\nWriting Slurm jobs using TensorFlow is as easy as creating the following scripts.
tf_script.py
#/usr/bin/env python\n\nimport tensorflow as tf\nprint(\"TensorFlow version:\", tf.__version__)\nprint(tf.config.list_physical_devices())\n\nmnist = tf.keras.datasets.mnist\n\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\nx_train, x_test = x_train / 255.0, x_test / 255.0\n\n\nmodel = tf.keras.models.Sequential([\n tf.keras.layers.Flatten(input_shape=(28, 28)),\n tf.keras.layers.Dense(128, activation='relu'),\n tf.keras.layers.Dropout(0.2),\n tf.keras.layers.Dense(10)\n])\n\npredictions = model(x_train[:1]).numpy()\nprint(predictions)\n\nprint(tf.nn.softmax(predictions).numpy())\n\n# ... and so on ;-)\ntf_job.sh
#!/usr/bin/bash\n\n#SBATCH --job-name=tf-job\n#SBATCH --mem=10g\n#SBATCH --partition=gpu\n#SBATCH --gres=gpu:tesla:1\n\nsource $HOME/work/miniforge/bin/activate\nconda activate python-tf\n\npython tf_script.py &>tf-out.txt\nAnd then calling
$ sbatch tf_job.sh\nYou can find the reuslts in tf-out.txt after completion.
$ cat tf-out.txt \n2022-03-09 18:05:54.628846: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: SSE4.1 SSE4.2 AVX AVX2 AVX512F FMA\nTo enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.\n2022-03-09 18:05:56.999848: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1510] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 30988 MB memory: -> device: 0, name: Tesla V100-SXM2-32GB, pci bus id: 0000:18:00.0, compute capability: 7.0\nTensorFlow version: 2.6.2\n[PhysicalDevice(name='/physical_device:CPU:0', device_type='CPU'), PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]\n[[-0.07757086 0.04676083 0.9420195 -0.59902835 -0.26286742 -0.392514\n 0.3231195 -0.17169198 0.3480805 0.37013203]]\n[[0.07963609 0.09017922 0.22075593 0.04727634 0.06616627 0.05812084\n 0.11888511 0.07248258 0.12188996 0.12460768]]\nThis tutorial assumes familiarity with Linux/Unix operating systems. It also assumes that you have already connected to the cluster. We have collected some links to tutorials and manuals on the internet.
"},{"location":"hpc-tutorial/episode-0/#legend","title":"Legend","text":"Before we start with our first steps tutorial, we would like to introduce the following convention that we use throughout the series:
$ Commands are prefixed with a little dollar sign\nWhile file paths are highlighted like this: /data/cephfs-1/work/projects/cubit/current.
After connecting to the cluster, you are located on a login node. To get to your first compute node, type srun --time 7-00 --mem=8G --cpus-per-task=8 --pty bash -i which will launch an interactive Bash session on a free remote node running up to 7 days, enabling you to use 8 cores and 8 Gb memory. Typing exit will you bring back to the login node.
hpc-login-1$ srun -p long --time 7-00 --mem=8G --cpus-per-task=8 --pty bash -i\nhpc-cpu-1$ exit\n$\nSee? That was easy!
"},{"location":"hpc-tutorial/episode-0/#preparation","title":"Preparation","text":"In preparation for our first steps tutorial series, we would like you to install the software for this tutorial. In general the users on the cluster will manage their own software with the help of conda. If you haven't done so so far, please follow the instructions in installing conda first. The only premise is that you are able to log into the cluster. Make also sure that you are logged in to a computation node using srun -p medium --time 1-00 --mem=4G --cpus-per-task=1 --pty bash -i.
Now we will create a new environment, so as to not interfere with your current or planned software stack, and install into it all the software that we need during the tutorial. Run the following commands:
$ conda create -n first-steps python=3 snakemake bwa delly samtools gatk4\n$ conda activate first-steps\n(first-steps) $\nThis is part one of the \"First Steps\" BIH Cluster Tutorial. Here we will build a small pipeline with alignment and variant calling. The premise is that you have the tools installed as described in Episode 0. For this episode, please make sure that you are on a compute node. As a reminder, the command to access a compute node with the required resources is
$ srun --time 7-00 --mem=8G --cpus-per-task=8 --pty bash -i\nWe will provide you with some example FASTQ files, but you can use your own if you like. You can find the data here:
/data/cephfs-1/work/projects/cubit/tutorial/input/test_R1.fq.gz/data/cephfs-1/work/projects/cubit/tutorial/input/test_R2.fq.gzFirst, you should create a folder where the output of this tutorial will go. It would be good to have it in your work directory in /data/cephfs-1/home/users/$USER, because it is faster and there is more space available.
(first-steps) $ mkdir -p /data/cephfs-1/home/users/$USER/work/tutorial/episode1\n(first-steps) $ pushd /data/cephfs-1/home/users/$USER/work/tutorial/episode1\nQuotas / File System limits
/data/cephfs-1/home/users/$USER. The reason for this is that nightly snapshots and backups are created for this directory which are precious resources./data/cephfs-1/home/users/$USER/work. The limits are much higher here but no snapshots or backups are available./data/cephfs-1/home/users/$USER/scratch. However, files placed here are automatically removed after 2 weeks. This is only appropriate for files during download or temporary files.In general it is advisable to have a proper temporary directory available. You can create one in your ~/scratch folder and make it available to the system.
(first-steps) $ export TMPDIR=/data/cephfs-1/home/users/$USER/scratch/tmp\n(first-steps) $ mkdir -p $TMPDIR\nThe static data is located in /data/cephfs-1/work/projects/cubit/current/static_data. For our small example, the required reference genome and index can be found at:
/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta/data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fastaLet's align our data:
(first-steps) $ bwa mem -t 8 \\\n -R \"@RG\\tID:FLOWCELL.LANE\\tPL:ILLUMINA\\tLB:test\\tSM:PA01\" \\\n /data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fasta \\\n /data/cephfs-1/work/projects/cubit/tutorial/input/test_R1.fq.gz \\\n /data/cephfs-1/work/projects/cubit/tutorial/input/test_R2.fq.gz \\\n| samtools view -b \\\n| samtools sort -O BAM -T $TMPDIR -o aln.bam\n\n(first-steps) $ samtools index aln.bam\nAnd do the structural variant calling:
(first-steps) $ delly call \\\n -g /data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta \\\n aln.bam\nNote that delly will not find any variants.
"},{"location":"hpc-tutorial/episode-1/#small-variant-calling-snv-indel","title":"Small Variant Calling (SNV, indel)","text":"And now for the SNP calling (this step will take ~ 20 minutes):
(first-steps) $ gatk HaplotypeCaller \\\n -R /data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta \\\n -I aln.bam \\\n -ploidy 2 \\\n -O test.GATK.vcf\nSo this is it! We used the tools that we installed previously, accessed the reference data and ran a simple alignment and variant calling pipeline. You can access a list of all static data through this wiki, follow this link to the Static Data. You can also have a peek via:
(first-steps) $ tree -L 3 /data/cephfs-1/work/projects/cubit/current/static_data | less\nWelcome to the second episode of our tutorial series!
Once you are logged in to the cluster, you have the possibility to distribute your jobs to all the nodes that are available. But how can you do this easily? The key command to this magic is sbatch. This tutorial will show you how you can use this efficiently.
sbatch Command","text":"So what is sbatch doing for you?
You use the sbatch command in front of the script you actually want to run. sbatch then puts your job into the job queue. The job scheduler looks at the current status of the whole system and will assign the first job in the queue to a node that is free in terms of computational load. If all machines are busy, yours will wait. But your job will sooner or later get assigned to a free node.
We strongly recommend using this process for starting your computationally intensive tasks because you will get the best performance for your job and the whole system won't be disturbed by jobs that are locally blocking nodes. Thus, everybody using the cluster benefits.
You may have noticed that you run sbatch with a script, not with regular commands. The reason is that sbatch only accepts bash scripts. If you give sbatch a normal shell command or binary, it won't work. This means that we have to put the command(s) we want to use in a bash script. A skeleton script can be found at /data/cephfs-1/work/projects/cubit/tutorial/skeletons/submit_job.sh
The content of the file:
#!/bin/bash\n\n# Set a name for the job (-J or --job-name).\n#SBATCH --job-name=tutorial\n\n# Set the file to write the stdout and stderr to (if -e is not set; -o or --output).\n#SBATCH --output=logs/%x-%j.log\n\n# Set the number of cores (-c or --cpus-per-task).\n#SBATCH --cpus-per-task=8\n\n# Force allocation of the two cores on ONE node.\n#SBATCH --nodes=1\n\n# Set the total memory. Units can be given in T|G|M|K.\n#SBATCH --mem=8G\n\n# Optionally, set the partition to be used (-p or --partition).\n#SBATCH --partition=medium\n\n# Set the expected running time of your job (-t or --time).\n# Formats are MM:SS, HH:MM:SS, Days-HH, Days-HH:MM, Days-HH:MM:SS\n#SBATCH --time=30:00\n\nexport TMPDIR=/data/cephfs-1/home/users/${USER}/scratch/tmp\nmkdir -p ${TMPDIR}\nThe lines starting with #SBATCH are actually setting parameters for a sbatch command, so #SBATCH --job-name=tutorial is equal to sbatch --job-name=tutorial. Slurm will create a log file with a file name composed of the job name (%x) and the job ID (%j), e.g. logs/tutorial-XXXX.log. It will not automatically create the logs directory, we need to do this manually first. Here, we emphasize the importance of the log files! They are the first place to look if anything goes wrong.
To start now with our tutorial, create a new tutorial directory with a log directory, e.g.,
(first-steps) $ mkdir -p /data/cephfs-1/home/users/$USER/work/tutorial/episode2/logs\nand copy the wrapper script to this directory:
(first-steps) $ pushd /data/cephfs-1/home/users/$USER/work/tutorial/episode2\n(first-steps) $ cp /data/cephfs-1/work/projects/cubit/tutorial/skeletons/submit_job.sh .\n(first-steps) $ chmod u+w submit_job.sh\nNow open this file and copy the same commands we executed in the last tutorial to this file.
To keep it simple, we will put everything into one script. This is perfectly fine because the alignment and indexing are sequential. But there are two steps that could be run in parallel, namely the variant calling, because they don't depend on each other. We will learn how to do that in a later tutorial. Your file should look something like this:
#!/bin/bash\n\n# Set a name for the job (-J or --job-name).\n#SBATCH --job-name=tutorial\n\n# Set the file to write the stdout and stderr to (if -e is not set; -o or --output).\n#SBATCH --output=logs/%x-%j.log\n\n# Set the number of cores (-c or --cpus-per-task).\n#SBATCH --cpus-per-task=8\n\n# Force allocation of the two cores on ONE node.\n#SBATCH --nodes=1\n\n# Set the total memory. Units can be given in T|G|M|K.\n#SBATCH --mem=8G\n\n# Optionally, set the partition to be used (-p or --partition).\n#SBATCH --partition=medium\n\n# Set the expected running time of your job (-t or --time).\n# Formats are MM:SS, HH:MM:SS, Days-HH, Days-HH:MM, Days-HH:MM:SS\n#SBATCH --time=30:00\n\nexport TMPDIR=/data/cephfs-1/home/users/${USER}/scratch/tmp\nmkdir -p ${TMPDIR}\n\nBWAREF=/data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fasta\nREF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\nbwa mem -t 8 \\\n -R \"@RG\\tID:FLOWCELL.LANE\\tPL:ILLUMINA\\tLB:test\\tSM:PA01\" \\\n $BWAREF \\\n /data/cephfs-1/work/projects/cubit/tutorial/input/test_R1.fq.gz \\\n /data/cephfs-1/work/projects/cubit/tutorial/input/test_R2.fq.gz \\\n| samtools view -b \\\n| samtools sort -O BAM -T $TMPDIR -o aln.bam\n\nsamtools index aln.bam\n\ndelly call -g \\\n $REF \\\n aln.bam\n\ngatk HaplotypeCaller \\\n -R $REF \\\n -I aln.bam \\\n -ploidy 2 \\\n -O test.GATK.vcf\nLet's run it (make sure that you are in the tutorial/episode2 directory!):
(first-steps) $ sbatch submit_job.sh\nAnd wait for the response which will tell you that your job was submitted and which job id number it was assigned. Note that sbatch only tells you that the job has started, but nothing about finishing. You won't get any response at the terminal when the job finishes. It will take approximately 20 minutes to finish the job.
You'll probably want to see how your job is doing. You can get a list of your jobs using:
(first-steps) $ squeue --me\nNote that logins are also considered as jobs.
Identify your job by the <JOBID> (1st column) or the name of the script (3rd column). The most likely states you will see (5th column of the table):
PD pending, waiting to be submittedR runningIn the 8th column you can see that your job is very likely running on a different machine than the one you are on!
Do not use Slurm and watch or loops
The watch command is a useful tool for running commands in a loop every N seconds. For example, on your workstation you could do watch 'ping -c 3 google.com' to execute three network pings to Google every two seconds.
\ud83d\udc4e Using watch or manual loops in a cluster environment can have bad effects when querying Slurm or the shared file system. Both are shared resources and \"expensive\" queries should not be run in loops. For Slurm, this includes running squeue. The same would be true for running squeue -i which performs an internal loop.
\ud83d\udc4d Use the Slurm query commands only when you actually need the output. If you run them in an (implict or explicit) loop, then do so only for a short time and don't leave this open in a screen.
Get more information about your jobs by either passing the job id:
(first-steps) $ sstat <JOBID>\nAnd of course, watch what the logs are telling you:
(first-steps) $ tail -f logs/tutorial-<JOBID>.log\nThere will be no notification when your job is done, so it is best to watch the squeue --me command. To watch the sbatch command there is a linux command watch that you give a command to execute every few seconds. This is useful for looking for changes in the output of a command. The seconds between two executions can be set with the -n option. It is best to use -n 60 to minimize unnecessary load on the file system:
(first-steps) $ watch -n 60 squeue --me\nscancel with your job-ID: (first-steps) $ scancel <job-ID>\nThe cluster has a special way of organizing itself and by telling the cluster how long and with which priority you want your jobs to run, you can help it in this. There is a system set up on the cluster where you can enqueue your jobs to so-called partitions. partitions have different prioritites and are allowed for different running times. To get to know what partitions are available, and how to use them properly, we highly encourage you to read the cluster queues wiki page.
"},{"location":"hpc-tutorial/episode-3/","title":"First Steps: Episode 3","text":"Episode Topic 0 How can I install the tools? 1 How can I use the static data? 2 How can I distribute my jobs on the cluster (Slurm)? 3 How can I organize my jobs with Snakemake? 4 How can I combine Snakemake and Slurm?In this episode we will discuss how we can parallelize steps in a pipeline that are not dependent on each other. In the last episode we saw a case (the variant calling) that could have been potentially parallelized.
We will take care of that today. Please note that we are not going to use the sbatch command we learned earlier. Thus, this tutorial will run on the same node where you execute the script. We will introduce you to Snakemake, a tool with which we can model dependencies and run things in parallel. In the next tutorial we will learn how to submit the jobs with sbatch and Snakemake combined.
For those who know make already, Snakemake will be familiar. You can think of Snakemake being a bunch of dedicated bash scripts that you can make dependent on each other. Snakemake will start the next script when a previous one finishes, and potentially it will run things in parallel if the dependencies allow.
Snakemake can get confusing, especially if the project gets big. This tutorial will only cover the very basics of this powerful tool. For more, we highly recommend digging into the Snakemake documentation:
Every Snakemake run requires a Snakefile file. Create a new folder inside your tutorial folder and copy the skeleton:
(first-steps) $ mkdir -p /data/cephfs-1/home/users/${USER}/work/tutorial/episode3\n(first-steps) $ pushd /data/cephfs-1/home/users/${USER}/work/tutorial/episode3\n(first-steps) $ cp /data/cephfs-1/work/projects/cubit/tutorial/skeletons/Snakefile .\n(first-steps) $ chmod u+w Snakefile\nYour Snakefile should look as follows:
rule all:\n input:\n 'snps/test.vcf',\n 'structural_variants/test.vcf'\n\nrule alignment:\n input:\n '/data/cephfs-1/work/projects/cubit/tutorial/input/test_R1.fq.gz',\n '/data/cephfs-1/work/projects/cubit/tutorial/input/test_R2.fq.gz',\n output:\n bam='alignment/test.bam',\n bai='alignment/test.bam.bai',\n shell:\n r\"\"\"\n export TMPDIR=/data/cephfs-1/home/users/${{USER}}/scratch/tmp\n mkdir -p ${{TMPDIR}}\n\n BWAREF=/data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n bwa mem -t 8 \\\n -R \"@RG\\tID:FLOWCELL.LANE\\tPL:ILLUMINA\\tLB:test\\tSM:PA01\" \\\n ${{BWAREF}} \\\n {input} \\\n | samtools view -b \\\n | samtools sort -O BAM -T ${{TMPDIR}} -o {output.bam}\n\n samtools index {output.bam}\n \"\"\"\n\nrule structural_variants:\n input:\n 'alignment/test.bam'\n output:\n 'structural_variants/test.vcf'\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n delly call -o {output} -g ${{REF}} {input}\n \"\"\"\n\nrule snps:\n input:\n 'alignment/test.bam'\n output:\n 'snps/test.vcf'\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n gatk HaplotypeCaller \\\n -R ${{REF}} \\\n -I {input} \\\n -ploidy 2 \\\n -O {output}\n \"\"\"\nLet me explain. The content resembles the same steps we took in the previous tutorials. Although every step has its own rule (alignment, snp calling, structural variant calling), we could instead have written everything in one rule. It is up to you to design your rules! Note that the rule names are arbitrary and not mentioned anywhere else in the file.
But there is one primary rule: the rule all. This is the kickoff rule that makes everything run.
As you might have noticed, every rule has three main parameters: input, output and shell. input defines the files that are going into the rule, output those that are produced when executing the rule, and shell is the bash script that processes input to produce output.
Rule all does not have any output or shell, it uses input to start the chain of rules. Note that the input files of this rule are the output files of rule snps and structural_variants. The input of those rules is the output of rule alignment. This is how Snakemake processes the rules: It looks for rule all (or a rule that just has input files) and figures out how it can create the required input files with other rules by looking at their output files (the input files of one rule must be the output files of another rule). In our case it traces the workflow back to rule snps and structural_variants as they have the matching output files. They depend in return on the alignment, so the alignment rule must be executed, and this is the first thing that will be done by Snakemake.
There are also some peculiarities about Snakemake:
input or output as is done in rule alignment with the output files.input and output files in the script by writing {input} or {output}.{output.bam}${{VAR}} instead of ${VAR} but not Snakemake internal variables like {input} or {output}structural_variants we cheat a bit because delly does not produce output files if it can't find variants.touching (i.e., creating) the required output file. Snakemake has a function for doing so (call touch() on the filename).But Snakemake can do more. It is able to parse the paths of the output files and set wildcards if you want. For this your input (and output) file names have to follow a parsable scheme. In our case they do! Our FASTQ files, our only initial input files, start with test. The output of the alignment as well as the variant calling is also prefixed test. We now can modify the Snakemake file accordingly, by exchanging every occurrence of test in each input or output field with {id} (note that you could also give a different name for your variable). Only the input rule should not be touched, otherwise Snakemake would not know which value this variable should have. Your Snakefile should look now like this:
rule all:\n input:\n 'snps/test.vcf',\n 'structural_variants/test.vcf'\n\nrule alignment:\n input:\n '/data/cephfs-1/work/projects/cubit/tutorial/input/{id}_R1.fq.gz',\n '/data/cephfs-1/work/projects/cubit/tutorial/input/{id}_R2.fq.gz',\n output:\n bam='alignment/{id}.bam',\n bai='alignment/{id}.bam.bai',\n shell:\n r\"\"\"\n export TMPDIR=/data/cephfs-1/home/users/${{USER}}/scratch/tmp\n mkdir -p ${{TMPDIR}}\n\n BWAREF=/data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n bwa mem -t 8 \\\n -R \"@RG\\tID:FLOWCELL.LANE\\tPL:ILLUMINA\\tLB:test\\tSM:PA01\" \\\n ${{BWAREF}} \\\n {input} \\\n | samtools view -b \\\n | samtools sort -O BAM -T ${{TMPDIR}} -o {output.bam}\n\n samtools index {output.bam}\n \"\"\"\n\nrule structural_variants:\n input:\n 'alignment/{id}.bam'\n output:\n 'structural_variants/{id}.vcf'\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n delly call -o {output} -g ${{REF}} {input}\n \"\"\"\n\nrule snps:\n input:\n 'alignment/{id}.bam'\n output:\n 'snps/{id}.vcf'\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n gatk HaplotypeCaller \\\n -R ${{REF}} \\\n -I {input} \\\n -ploidy 2 \\\n -O {output}\n \"\"\"\nBefore we finally run this, we can make a dry run. Snakemake will show you what it would do:
(first-steps) $ snakemake -n\nIf everything looks green, you can run it for real. We provide it two cores to allow two single-threaded jobs to be run simultaneously:
(first-steps) $ snakemake -j 2\nIn the last episodes we learned about distributing a job among the cluster nodes using sbatch and how to automate and parallelize our pipeline with Snakemake. We are lucky that those two powerful commands can be combined. What is the result? You will have an automated pipeline with Snakemake that uses sbatch to distribute jobs among the cluster nodes instead of running only the same node.
The best thing is that we can reuse our Snakefile as it is and just write a wrapper script to call Snakemake. We run the script and the magic will start.
First, create a new folder for this episode:
(first-steps) $ mkdir -p /data/cephfs-1/home/users/${USER}/work/tutorial/episode4/logs\n(first-steps) $ pushd /data/cephfs-1/home/users/${USER}/work/tutorial/episode4\nAnd copy the wrapper script to this folder as well as the Snakefile (you can also reuse the one with the adjustments from the previous episode):
(first-steps) $ cp /data/cephfs-1/work/projects/cubit/tutorial/skeletons/submit_snakejob.sh .\n(first-steps) $ cp /data/cephfs-1/work/projects/cubit/tutorial/skeletons/Snakefile .\n(first-steps) $ chmod u+w submit_snakejob.sh Snakefile\nThe Snakefile is already known to you but let me explain the wrapper script submit_snakejob.sh:
#!/bin/bash\n\n# Set a name for the job (-J or --job-name).\n#SBATCH --job-name=tutorial\n\n# Set the file to write the stdout and stderr to (if -e is not set; -o or --output).\n#SBATCH --output=logs/%x-%j.log\n\n# Set the number of cores (-c or --cpus-per-task).\n#SBATCH --cpus-per-task=2\n\n# Force allocation of the two cores on ONE node.\n#SBATCH --nodes=1\n\n# Set the total memory. Units can be given in T|G|M|K.\n#SBATCH --mem=1G\n\n# Optionally, set the partition to be used (-p or --partition).\n#SBATCH --partition=medium\n\n# Set the expected running time of your job (-t or --time).\n# Formats are MM:SS, HH:MM:SS, Days-HH, Days-HH:MM, Days-HH:MM:SS\n#SBATCH --time=30:00\n\n\nexport TMPDIR=/data/cephfs-1/home/users/${USER}/scratch/tmp\nexport LOGDIR=logs/${SLURM_JOB_NAME}-${SLURM_JOB_ID}\nmkdir -p $LOGDIR\n\neval \"$($(which conda) shell.bash hook)\"\nconda activate first-steps\n\nset -x\n\nsnakemake --profile=cubi-v1 -j 2 -k -p --restart-times=2\nIn the beginning you see the #SBATCH that introduces the parameters when you provide this script to sbatch as described in the second episode. Please make sure that the logs folder exists before starting the run! We then set and export the TMPDIR and LOGDIR variables. Note that LOGDIR has a subfolder named $SLURM_JOB_NAME-$SLURM_JOB_ID that will be created for you. Snakemake will store its logfiles for this very Snakemake run in this folder. The next new thing is set -x. This simply prints to the terminal every command that is executed within the script. This is useful for debugging.
Finally, the Snakemake call takes place. With the --profile option we define that Snakemake uses the Snakemake profile at /etc/xdg/snakemake/cubi-v1. The profile will take create appropriate calls to sbatch and interpret the following settings from your Snakemake rule:
threads: the number of threads to execute the job onk, M, G, or T. You can specify EITHERresources.mem/resources.mem_mb: the memory to allocate for the whole job, ORresources.mem_per_thread: the memory to allocate for each thread.resources.time: the running time of the rule, in a syntax supported by Slurm, e.g. HH:MM:SS or D-HH:MM:SSresources.partition: the partition to submit your job into (Slurm will pick a fitting partition for you by default)resources.nodes: the number of nodes to schedule your job on (defaults to 1 and you will want to keep that value unless you want to use MPI)The other options to snakemake have the meaning:
-j 2: run at most two jobs at the same time-k: keep going even if a rule execution fails-p: print the executed shell commands--restart-times=2: restart failing jobs up to two timesIt is now time to update your Snakefile such that it actually specifies the resources mentioned above:
rule all:\n input:\n 'snps/test.vcf',\n 'structural_variants/test.vcf'\n\nrule alignment:\n input:\n '/data/cephfs-1/work/projects/cubit/tutorial/input/{id}_R1.fq.gz',\n '/data/cephfs-1/work/projects/cubit/tutorial/input/{id}_R2.fq.gz',\n output:\n bam='alignment/{id}.bam',\n bai='alignment/{id}.bam.bai',\n threads: 8\n resources:\n mem='8G',\n time='12:00:00',\n shell:\n r\"\"\"\n export TMPDIR=/data/cephfs-1/home/users/${{USER}}/scratch/tmp\n mkdir -p ${{TMPDIR}}\n\n BWAREF=/data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n bwa mem -t 8 \\\n -R \"@RG\\tID:FLOWCELL.LANE\\tPL:ILLUMINA\\tLB:test\\tSM:PA01\" \\\n ${{BWAREF}} \\\n {input} \\\n | samtools view -b \\\n | samtools sort -O BAM -T ${{TMPDIR}} -o {output.bam}\n\n samtools index {output.bam}\n \"\"\"\n\nrule structural_variants:\n input:\n 'alignment/{id}.bam'\n output:\n 'structural_variants/{id}.vcf'\n threads: 1\n resources:\n mem='4G',\n time='2-00:00:00',\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n delly call -o {output} -g ${{REF}} {input}\n \"\"\"\n\ndef snps_mem(wildcards, attempt):\n mem = 2 * attempt\n return '%dG' % mem\n\nrule snps:\n input:\n 'alignment/{id}.bam'\n output:\n 'snps/{id}.vcf'\n threads: 1\n resources:\n mem=snps_mem,\n time='04:00:00',\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n gatk HaplotypeCaller \\\n -R ${{REF}} \\\n -I {input} \\\n -ploidy 2 \\\n -O {output}\n \"\"\"\nWe thus configure the resource consumption of the rules as follows:
alignment with 8 threads and up to 8GB of memory in total with a running time of up to 12 hours,structural_variants with one thread and up to 4GB of memory in with a running time of up to 2 days,snps with one thread and running up to four hours. Instead of passing a static amount of memory, we pass a resource callable. The attempt parameter will be passed a value of 1 on the initial invocation. If variant calling with the GATK HaplotypeCaller fails then it will retry and attempt will have an incremented value on each invocation (2 on the first retry and so on). Thus, we try to do small variant calling with 2, 4, 6, and 8 GB.Finally, run the script:
(first-steps) $ sbatch submit_snakejob.sh\nIf you watch squeue --me now, you will see that the jobs are distributed to the system:
(first-steps) $ squeue --me\nPlease refer to the Snakemake documentation for more details on using Snakemake, in particular how to use the cluster configuration on how to specify the resource requirements on a per-rule base.
"},{"location":"misc/external-resources/","title":"External Resources","text":""},{"location":"misc/external-resources/#basic-linux","title":"Basic Linux","text":"The BIH HPC uses CentOS Linux. A basic understanding of Linux is required. Even better, you should already have intermediate to advanced Linux/Unix skills.
BIH HPC IT cannot provide you with basic Unix training. Please ask your home organization (e.g., Charite or MDC) to provide you with basic Linux training.
That said, here are some resources that we find useful:
"},{"location":"misc/external-resources/#internet-tutorials","title":"Internet Tutorials","text":"There is a large number of Linux tutorials online including:
GOBLET has a number of Bioinformatics-focused tutorials. This includes
Some software is provided by HPC Administration based on the criteria that it is:
Currently, this includes:
On the GPU node, this also includes a recent NVIDIA CUDA version.
To see which software is available, use module avail on a compute node (this will not work on login nodes):
$ module avail\n--------------------- /opt/local/modules ---------------------\ncmake/3.11.0-0 llvm/6.0.0-0\ngcc/7.2.0-0 openmpi/4.0.3-0\nTo load software, use module load. This will adjust the environment variables accordingly, in particular update PATH such that the executable are available.
$ which gcc\n/bin/gcc\n$ module load gcc/7.2.0-0\n$ which gcc\n/opt/local/gcc-7.2.0-0/bin/gcc\nProblems with executing module?
See the corresponding FAQ entry in the case that you get a -bash: module: command not found when calling module.
The BIH Cluster is a valuable resource. It has been used to support the publications listed below.
Chopra A, Song J, Weiner J, Beule D, Schaefer AS (2024). \u201cGenetic analysis of cis-enhancers associated with bone mineral density and periodontitis in the gene SOST.\u201d ResearchSquare. \u261e Link
Opitz B, Essex M, Pascual-Leone BM, Loeber U, Kuhring M, Zhang B, Bruening U, Fritsche-Guenther R, Krzanowski M, Vernengo FF, Brumhard S, R\u00f6wekamp I, Bielecka A, Lesker T, Wyler E, Landthaler M, Mantei A, Meisel C, Caesar S, Thibeault C, Corman V, Marko L, Suttorp N, Strowig T, Kurth F, Sander L, Li Y, Kirwan J, Forslund S (2023). \u201cGut microbiota dysbiosis is associated with altered tryptophan metabolism and dysregulated inflammatory response in severe COVID-19.\u201d doi:10.21203/rs.3.rs-2459463/v1 \u261e Link
Zuljan E, von der Emde B, Piwonski I, Pestana ACA, Klinghammer K, Mock A, Horak P, Heining C, Klauschen F, Pretzell M, Boerries I, Brandts C, Kreutzfeldt S, Teleanu M, H\u00fcbschmann D, Morris L, Keller U, Glimm H, Fr\u00f6hling S, Ochsenreither S, Keilholz U, Blanc E, Beule D, Rieke D (2024). \u201cIntegrative analysis reveals a macrophage-predominant, immunosuppressive immune microenvironment and subtype-specific therapeutic vulnerabilities in advanced salivary gland cancer.\u201d medRxiv, 2024-06. doi:10.1101/2024.06.11.24308538 \u261e Link
Yerinde C, Keye J, Durlanik S, Freise I, Nowak F, Hsiao H, Letizia M, Schlickeiser S, Obermayer B, Huck A, Friedrich M, Wu H, Kunkel D, K\u00fchl AA, Bauer S, Thiel A, Siegmund B, Glauben R, Weidinger C (2022). \u201cHDAC7 controls anti-viral and anti-tumor immunity by CD8+ T cells.\u201d bioRxiv. doi:10.1101/2022.09.18.508452 \u261e Link
Winkler W, D\u00edaz C, Blanc E, Napieczynska H (2021). \u201cPreclinical models of human multiple myeloma subgroups.\u201d bioRxiv. doi:10.1101/2021.08.28.458010 \u261e Link
G\u00f6ppel T, Obermayer B, Chen I, Gerland U (2021). \u201cA kinetic error filtering mechanism for enzyme-free copying of nucleic acid sequences.\u201d bioRxiv. doi:10.1101/2021.08.06.455386 \u261e Link
"},{"location":"misc/publication-list/#2024","title":"2024","text":"K\u00f6hler RM, Binns TS, Merk T, Zhu G, Yin Z, Zhao B, Chikermane M, Vanhoecke J, Busch JL, Habets JGV, Faust K, Schneider GH, Cavallo A, Haufe S, Zhang J, K\u00fchn AA, Haynes JD, Neumann WJ. Dopamine and deep brain stimulation accelerate the neural dynamics of volitional action in Parkinson\u2019s disease. Brain. 2024 Oct 1;147(10):3358\u201369. doi: 10.1093/brain/awae219
Binns TS, K\u00f6hler RM, Vanhoecke J, Chikermane M, Gerster M, Merk T, Pellegrini F, Busch JL, Habets JGV, Cavallo A, Beyer JC, Al-Fatly B, Li N, Horn A, Krause P, Faust K, Schneider GH, Haufe S, K\u00fchn AA, Neumann WJ. Shared pathway-specific network mechanisms of dopamine and deep brain stimulation for the treatment of Parkinson\u2019s disease. bioRxiv. 2024; p. 2024.04.14.586969. Available from: doi: 10.1101/2024.04.14.586969v2
Hollunder, B., Ostrem, J.L., Sahin, I.A., Rajamani, N., Oxenford, S., Butenko, K., Neudorfer, C., Reinhardt, P., Zvarova, P., Polosan, M., Akram, H., Vissani, M., Zhang, C., Sun, B., Navratil, P., Reich, M.M., Volkmann, J., Yeh, F.-C., Baldermann, J.C., Dembek, T.A., Visser-Vandewalle, V., Alho, E.J.L., Franceschini, P.R., Nanda, P., Finke, C., K\u00fchn, A.A., Dougherty, D.D., Richardson, R.M., Bergman, H., DeLong, M.R., Mazzoni, A., Romito, L.M., Tyagi, H., Zrinzo, L., Joyce, E.M., Chabardes, S., Starr, P.A., Li, N., Horn, A., 2024. Mapping dysfunctional circuits in the frontal cortex using deep brain stimulation. Nat. Neurosci. 1\u201314. doi: 10.1038/s41593-024-01570-1
Schmidt A, Danyel M, Grundmann K, Brunet T, Klinkhammer H, Hsieh T, Engels H, Peters S, Knaus A, Moosa S, Averdunk L, Boschann F, Sczakiel HL, Schwartzmann S, Mensah MA, Pantel JT, Holtgrewe M, B\u00f6sch A, Wei\u00df C, Weinhold N, Suter A, Stoltenburg C, Neugebauer J, Kallinich T, Kaindl AM, Holzhauer S, B\u00fchrer C, Bufler P, Kornak U, Ott C, Sch\u00fclke M, Nguyen HHP, Hoffjan S, Grasemann C, Rothoeft T, Brinkmann F, Matar N, Sivalingam S, Perne C, Mangold E, Kreiss M, Cremer K, Betz RC, M\u00fccke M, Grigull L, Klockgether T, Spier I, Heimbach A, Bender T, Brand F, Stieber C, Morawiec AM, Karakostas P, Sch\u00e4fer VS, Bernsen S, Weydt P, Castro-Gomez S, Aziz A, Grobe-Einsler M, Kimmich O, Kobeleva X, \u00d6nder D, Lesmann H, Kumar S, Tacik P, Basin MA, Incardona P, Lee-Kirsch MA, Berner R, Schuetz C, K\u00f6rholz J, Kretschmer T, Donato ND, Schr\u00f6ck E, Heinen A, Reuner U, Han\u00dfke A, Kaiser FJ, Manka E, Munteanu M, Kuechler A, Cordula K, Hirtz R, Schlapakow E, Schlein C, Lisfeld J, Kubisch C, Herget T, Hempel M, Weiler-Normann C, Ullrich K, Schramm C, Rudolph C, Rillig F, Groffmann M, Muntau A, Tibelius A, Schwaibold EMC, Schaaf CP, Zawada M, Kaufmann L, Hinderhofer K, Okun PM, Kotzaeridou U, Hoffmann GF, Choukair D, Bettendorf M, Spielmann M, Ripke A, Pauly M, M\u00fcnchau A, Lohmann K, H\u00fcning I, Hanker B, B\u00e4umer T, Herzog R, Hellenbroich Y, Westphal DS, Strom T, Kovacs R, Riedhammer KM, Mayerhanser K, Graf E, Brugger M, Hoefele J, Oexle K, Mirza-Schreiber N, Berutti R, Schatz U, Krenn M, Makowski C, Weigand H, Schr\u00f6der S, Rohlfs M, Vill K, Hauck F, Borggraefe I, M\u00fcller-Felber W, Kurth I, Elbracht M, Knopp C, Begemann M, Kraft F, Lemke JR, Hentschel J, Platzer K, Strehlow V, Jamra RA, Kehrer M, Demidov G, Beck-W\u00f6dl S, Graessner H, Sturm M, Zeltner L, Sch\u00f6ls LJ, Magg J, Bevot A, Kehrer C, Kaiser N, Turro E, Horn D, Gr\u00fcters-Kieslich A, Klein C, Mundlos S, N\u00f6then M, Riess O, Meitinger T, Krude H, Krawitz PM, Haack T, Ehmke N, Wagner M (2024). \u201cNext-generation phenotyping integrated in a national framework for patients with ultrarare disorders improves genetic diagnostics and yields new molecular findings.\u201d Nature Genetics. ISSN 1546-1718, doi:10.1038/s41588-024-01836-1 \u261e Link
Marot-Lassauzaie V, Beneyto-Calabuig S, Obermayer B, Velten L, Beule D, Haghverdi L (2024). \u201cIdentifying cancer cells from calling single-nucleotide variants in scRNA-seq data.\u201d Bioinformatics, 40(9), btae512. ISSN 1367-4811, doi:10.1093/bioinformatics/btae512 \u261e Link
Horn V, Cancino CA, Steinheuer LM, Obermayer B, Fritz K, Nguyen AL, Juhran KS, Plattner C, B\u00f6sel D, Oldenburg L, Burns M, Schulz AR, Saliutina M, Mantzivi E, Lissner D, Conrad T, Mashreghi M, Zundler S, Sonnenberg E, Schumann M, Haag L, Beule D, Flatz L, Atreya I, Atreya R, Bacher P, Becker C, Bojarski C, Britzen-Laurent N, Bosch-Voskens C, Chang H, Diefenbach A, G\u00fcnther C, Hegazy AN, Hildner K, Klose CS, Koop K, Krug S, K\u00fchl AA, Leppkes M, L\u00f3pez-Posadas R, Ludwig LS, Neufert C, Neurath M, Patankar J, Pr\u00fc\u00df M, Radbruch A, Romagnani C, Ronchi F, Sanders A, Scheffold A, Schulzke J, Schumann M, Sch\u00fcrmann S, Siegmund B, St\u00fcrzl M, Trajanoski Z, Triantafyllopoulou A, Waldner M, Weidinger C, Wirtz S, Zundler S, Trjanoski Z, D\u2019Haens G, Weidinger C, Mei HE, Siegmund B, Thurley K, Hegazy AN (2024). \u201cMultimodal Profiling of Peripheral Blood Identifies Proliferating Circulating Effector CD4+ T Cells as Predictors for Response to Integrin \u03b14\u03b27\u2013Blocking Therapy in Inflammatory Bowel Disease.\u201d Gastroenterology. ISSN 0016-5085, doi:10.1053/j.gastro.2024.09.021 \u261e Link
Menzel M, Martis-Thiele M, Goldschmid H, Ott A, Romanovsky E, Siemanowski-Hrach J, Seillier L, Br\u00fcchle NO, Maurer A, Lehmann K, Begemann M, Elbracht M, Meyer R, Dintner S, Claus R, Meier-Kolthoff JP, Blanc E, M\u00f6bs M, Joosten M, Benary M, Basitta P, H\u00f6lscher F, Tischler V, Gro\u00df T, Kutz O, Prause R, William D, Horny K, Goering W, Sivalingam S, Borkhardt A, Blank C, Junk SV, Yasin L, Moskalev EA, Carta MG, Ferrazzi F, T\u00f6gel L, Wolter S, Adam E, Matysiak U, Rosenthal T, D\u00f6nitz J, Lehmann U, Schmidt G, Bartels S, Hofmann W, Hirsch S, Dikow N, G\u00f6bel K, Banan R, Hamelmann S, Fink A, Ball M, Neumann O, Rehker J, Kloth M, Murtagh J, Hartmann N, Jurmeister P, Mock A, Kumbrink J, Jung A, Mayr E, Jacob A, Trautmann M, Kirmse S, Falkenberg K, Ruckert C, Hirsch D, Immel A, Dietmaier W, Haack T, Marienfeld R, F\u00fcrstberger A, Niew\u00f6hner J, Gerstenmaier U, Eberhardt T, Greif PA, Appenzeller S, Maurus K, Doll J, Jelting Y, Jonigk D, M\u00e4rkl B, Beule D, Horst D, Wulf A, Aust D, Werner M, Reuter-Jessen K, Str\u00f6bel P, Auber B, Sahm F, Merkelbach-Bruse S, Siebolts U, Roth W, Lassmann S, Klauschen F, Gaisa NT, Weichert W, Evert M, Armeanu-Ebinger S, Ossowski S, Schroeder C, Schaaf CP, Malek N, Schirmacher P, Kazdal D, Pfarr N, Budczies J, Stenzinger A (2024). \u201cBenchmarking whole exome sequencing in the German network for personalized medicine.\u201d European Journal of Cancer, 211, 114306. ISSN 0959-8049, doi:10.1016/j.ejca.2024.114306 \u261e Link
Fatykhova D, Fritsch VN, Siebert K, Methling K, Lalk M, Busche T, Kalinowski J, Weiner J, Beule D, Bertrams W, others (2024). \u201cMicroenvironmental acidification by pneumococcal sugar consumption fosters barrier disruption and immune suppression in the human alveolus.\u201d European Respiratory Journal. doi:doi.org/10.1183/13993003.01983-2023 \u261e Link
Adhikary PP, Idowu T, Tan Z, Hoang C, Shanta S, Dumbani M, Mappalakayil L, Awasthi B, Bermudez M, Weiner J, Beule D, Wolber G, Page BD, Hedtrich S (2024). \u201cDisrupting TSLP-TSLP receptor interactions via putative small molecule inhibitors yields a novel and efficient treatment option for atopic diseases.\u201d EMBO Molecular Medicine, 1-27. doi:10.1038/s44321-024-00085-3 \u261e Link
Mattei D, Ivanov A, Hammer J, Ugursu B, Schalbetter S, Richetto J, Weber-Stadlbauer U, Mueller F, Scarborough J, Wolf SA, Kettenmann H, Wollscheid B, Beule D, Meyer U (2024). \u201cMicroglia undergo molecular and functional adaptations to dark and light phases in male laboratory mice.\u201d Brain, Behavior, and Immunity, 120, 571-583. ISSN 0889-1591, doi:10.1016/j.bbi.2024.07.007 \u261e Link
Kopp J, Koch LA, Lyubenova H, K\u00fcchler O, Holtgrewe M, Ivanov A, Dubourg C, Launay E, Brachs S, Mundlos S, Pellerin L, Mameli M, Paolicelli RC (2024). \u201cLoss-of-function variants affecting the STAGA complex component SUPT7L cause a developmental disorder with generalized lipodystrophy.\u201d Human Genetics, 143(5), 683-694. doi:10.1007/s00439-024-02669-y \u261e Link
Ruiz-Orera J, Miller DC, Greiner J, Genehr C, Grammatikaki A, Blachut S, Mbebi J, Patone G, Myronova A, Adami E, others (2024). \u201cEvolution of translational control and the emergence of genes and open reading frames in human and non-human primate hearts.\u201d Nature Cardiovascular Research, 1-19. doi:10.1038/s44161-024-00544-7 \u261e Link
Fisch A, Pestana A, Sachse V, Doll C, Hofmann E, Heiland M, Obermueller T, Heidemann J, Dommerich S, Schoppe D, Schallenberg S, Piwonski I, Blanc E, Tinhofer I (2024). \u201cFeasibility analysis of using patient-derived tumour organoids for treatment decision guidance in locally advanced head and neck squamous cell carcinoma.\u201d European journal of cancer (Oxford, England : 1990), 213, 115100. ISSN 1879-0852, doi:10.1016/j.ejca.2024.115100 \u261e Link
Al-Diab O, S\u00fcnkel C, Blanc E, Catar RA, Ashraf MI, Zhao H, Wang P, Rinschen MM, Fritsche-Guenther R, Grahammer F, Bachmann S, Beule D, Kirwan JA, Rajewsky N, Huber TB, G\u00fcrgen D, Kusch A (2024). \u201cSex-specific molecular signature of mouse podocytes in homeostasis and in response to pharmacological challenge with rapamycin.\u201d Biology of sex differences, 15, 72. ISSN 2042-6410, doi:10.1186/s13293-024-00647-7 \u261e Link
Trinks A, Milek M, Beule D, Kluge J, Florian S, Sers C, Horst D, Morkel M, Bischoff P (2024). \u201cRobust detection of clinically relevant features in single-cell RNA profiles of patient-matched fresh and formalin-fixed paraffin-embedded (FFPE) lung cancer tissue.\u201d Cellular Oncology, 47, 1-11. doi:10.5281/zenodo.7852154 \u261e Link
\u0160mid A, \u0160tajdohar M, Milek M, Urban\u010di\u010d D, Ku\u017eeli\u010dki NK, Tamm R, Metspalu A, Mlinari\u010d-Ra\u0161\u010dan I (2024). \u201cTranscriptome analysis reveals involvement of thiopurine S-methyltransferase in oxidation-reduction processes.\u201d European Journal of Pharmaceutical Sciences, 192, 106616. doi:10.1016/j.ejps.2023.106616 \u261e Link
D\u00f6rner PJ, Anandakumar H, R\u00f6wekamp I, Vernengo FF, Pascual-Leone BM, Krzanowski M, Sellmaier J, Br\u00fcning U, Fritsche-Guenther R, Pfannkuch L, Kurth F, Milek M, Igbokwe V, L\u00f6ber U, Gutbier B, Holstein M, Heinz GA, Mashreghi M, Schulte LN, Klatt A, Caesar S, Wienhold S, Offermanns S, Mack M, Witzenrath M, Jordan S, Beule D, Kirwan JA, Forslund SK, Wilck N, Bartolomaeus H, Heimesaat MM, Opitz B (2024). \u201cClinically used broad-spectrum antibiotics compromise inflammatory monocyte-dependent antibacterial defense in the lung.\u201d Nature Communications, 15(1), 2788.
R\u00f6wekamp I, Maschirow L, Rabes A, Vernengo FF, Hamann L, Heinz GA, Mashreghi M, Caesar S, Milek M, Fonseca ACF, Wienhold S, Nouailles G, Yao L, Mousavi S, Bruder D, Boehme JD, Puzianowska-Kuznicka M, Beule D, Witzenrath M, Group CS, L\u00f6hning M, Klose CSN, Heimesaat MM, Diefenbach A, Opitz B, Fuchs A, Engelmann M, Paul G, Ayoub M, Groehl K, Riedl K, Stolz D, Bauer W, Diehl-Wiesenecker EC, Terue IvW, Galtung N, Suttorp N, Witzenrath M, Wildberg C, Pley C, Zessin E, Schmager S, Schaaf B, Kremling J, Nickoleit-Bitzenberger D, Azzaui H, Hower M, Hempel F, Prebeg K, Popkirova K, Kolditz M, Schulte-Hubbert B, Langner S, Rohde G, Bellinghausen C, Gr\u03cbnewaldt A, Endres A, Frigerio C, Fiedler B, Panning M, Welte T, Pink I, Drick N, F\u03cbhner T, Klooster Mv, Steinberg T, Barten-Neiner G, Kr\u00f6ner W, Unruh O, Adaskina N, Eberhardt F, Julius C, Illig T, Klopp N, Pletz M, Schleenvoigt BT, Bahrs C, Moeser A, Ankert J, Sommerwerck U, Wintermantel T, Dr\u00f6mann D, Parschke P, Franzen K, Rupp J, Waldeck F, K\u00e4ding N, Spinner C, Erber J, Voit F, Schneider J, Falcone M, Tiseo G, Heigener D, Hering I, Albrich W, Rassouli F, Wirth B, Neurohr C, Essig A, Stenger S, Wallner M, Burgmann H, Traby L, Schubert L, nurses aas (2024). \u201cIL-33 controls IL-22-dependent antibacterial defense by modulating the microbiota.\u201d Proceedings of the National Academy of Sciences, 121(22), e2310864121. doi:10.1073/pnas.2310864121 \u261e Link
Wei T, Blanc E, Peidli S, Bischoff P, Trinks A, Horst D, Sers C, Bl\u00fcthgen N, Beule D, Morkel M, Obermayer B (2024). \u201cHigh-confidence calling of normal epithelial cells allows identification of a novel stem-like cell state in the colorectal cancer microenvironment.\u201d International Journal of Cancer, n/a(n/a). doi:10.1002/ijc.35079 \u261e Link
Hramyka D, Sczakiel HL, Zhao MX, Stolpe O, Nieminen M, Adam R, Danyel M, Einicke L, H\u00e4gerling R, Knaus A, Mundlos S, Schwartzmann S, Seelow D, Ehmke N, Mensah MA, Boschann F, Beule D, Holtgrewe M (2024). \u201cREEV: review, evaluate and explain variants.\u201d Nucleic Acids Research, gkae366. doi:10.1093/nar/gkae366 \u261e Link
Ramberger E, Sapozhnikova V, Ng YLD, Dolnik A, Ziehm M, Popp O, Str\"ang E, Kull M, Gr\"unschl\"ager F, Kr\"uger J, Benary M, M\"uller S, Gao X, Murgai A, Haji M, Schmidt A, Lutz R, Nogai A, Braune J, Laue D, Langer C, Khandanpour C, Bassermann F, D\"ohner H, Engelhardt M, Straka C, Hundemer M, Beule D, Haas S, Keller U, Einsele H, Bullinger L, Knop S, Mertins P, Kr\"onke J (2024). \u201cThe proteogenomic landscape of multiple myeloma reveals insights into disease biology and therapeutic opportunities.\u201d Nature Cancer. ISSN 2662-1347, doi:10.1038/s43018-024-00784-3 \u261e Link
Habringer S, Ihlow J, Kleo K, Klostermann A, Schmidt M, Chai L, Kn\u00f6dler M, Leyvraz S, Sigler C, Sinn B, Maschmeyer G, Jegodka Y, Benary M, Ott C, Tinhofer I, Sch\u00e4fer R, M\u00f6bs M, Keller U, Keilholz U, Rieke DT (2024). \u201cA diagnostic challenge of KIT p.V559D and BRAF p.G469A mutations in a paragastric mass.\u201d The Oncologist, oyae137. ISSN 1083-7159, doi:10.1093/oncolo/oyae137 \u261e Link
Shi W, Song J, Weiner J, Chopra A, Dommisch H, Beule D, Schaefer AS (2024). \u201clncRNA CDKN2B-AS1 regulates collagen expression.\u201d Human Genetics. doi:10.1007/s00439-024-02674-1 \u261e Link
Shi W, Song J, Dommisch H, Weiner J, Schaefer A (2024). \u201cSTAT1 binds to CAD and periodontitis risk SNP rs10757278, causing inhibition of lncRNA CDKN2B-AS1 expression, which activates transcription of collagen genes in cells of the extracellular matrix.\u201d In EUROPEAN JOURNAL OF HUMAN GENETICS, volume 32, 402-402. SPRINGERNATURE CAMPUS, 4 CRINAN ST, LONDON, N1 9XW, ENGLAND.
Zheng L, Chopra A, Weiner J, Dommisch H, Schaefer A (2024). \u201cSystematic search for microRNA regulated genes with roles in the etiology of periodontitis.\u201d In EUROPEAN JOURNAL OF HUMAN GENETICS, volume 32, 406-406. SPRINGERNATURE CAMPUS, 4 CRINAN ST, LONDON, N1 9XW, ENGLAND.
Lebedin M, Ratswohl C, Garg A, Schips M, Garc\u00eda CV, Spatt L, Thibeault C, Obermayer B, Weiner J, Vel\u00e1squez IM, others (2024). \u201cSoluble ACE2 correlates with severe COVID-19 and can impair antibody responses.\u201d iScience.
Tranter E, Frentsch M, H\u00fctter-Kr\u00f6nke ML, Vuong GL, Busch D, Loyal L, Henze L, Rosnev S, Blau I, Thiel A, Beule D, Bullinger L, Obermayer B, Na I (2024). \u201cComparable CD8+ T-cell responses to SARS-CoV-2 vaccination in single-cell transcriptomics of recently allogeneic transplanted patients and healthy individuals.\u201d Journal of Medical Virology, 96(3), e29539. doi:10.1002/jmv.29539 \u261e Link
Bhasin MA, Knaus A, Incardona P, Schmid A, Holtgrewe M, Elbracht M, Krawitz PM, Hsieh T (2024). \u201cEnhancing Variant Prioritization in VarFish through On-Premise Computational Facial Analysis.\u201d Genes, 15(3), 370.
Ermshaus A, Piechotta M, R\u00fcter G, Keilholz U, Leser U, Benary M (2024). \u201cpreon: Fast and accurate entity normalization for drug names and cancer types in precision oncology.\u201d Bioinformatics, 40(3), btae085. ISSN 1367-4811, doi:10.1093/bioinformatics/btae085 \u261e Link
Jahn D, Knapstein PR, Otto E, K\u00f6hli P, Sevecke J, Graef F, Graffmann C, Fuchs M, Jiang S, Rickert M, Erdmann C, Appelt J, Revend L, K\u00fcttner Q, Witte J, Rahmani A, Duda G, Xie W, Donat A, Schinke T, Ivanov A, Tchouto MN, Beule D, Frosch K, Baranowsky A, Tsitsilonis S, Keller J (2024). \u201cIncreased beta2-adrenergic signaling promotes fracture healing through callus neovascularization in mice.\u201d Science Translational Medicine, 16(743), eadk9129. doi:10.1126/scitranslmed.adk9129 \u261e Link
"},{"location":"misc/publication-list/#2023","title":"2023","text":"Monsorno K, Ginggen K, Ivanov A, Buckinx A, Lalive AL, Tchenio A, Benson S, Vendrell M, D\u2019Alessandro A, Beule D, others (2023). \u201cLoss of microglial MCT4 leads to defective synaptic pruning and anxiety-like behavior in mice.\u201d Nature Communications, 14(1), 5749. doi:10.1038/s41467-023-41502-4 \u261e Link
Markus D, Pelletier A, Boube M, Port F, Boutros M, Payre F, Obermayer B, Zanet J (2023). \u201cThe pleiotropic functions of Pri smORF peptides synchronize leg development regulators.\u201d PLOS Genetics, 19(10), 1-27. doi:10.1371/journal.pgen.1011004 \u261e Link
Ludwik KA, Opitz R, Jyrch S, Megges M, Weiner J, Beule D, K\u00fchnen P, Stachelscheid H (2023). \u201cGeneration of iPSC lines with SLC16A2: G410R or SLC16A2 knock out.\u201d Stem Cell Research, 103256. doi:10.1016/j.scr.2023.103256 \u261e Link
Schneemann M, Heils L, Moos V, Wei\u00df F, Krug SM, Weiner J, Beule D, Gerhard R, Schulzke J, B\u00fccker R (2023). \u201cA Colonic Organoid Model Challenged with the Large Toxins of Clostridioides difficile TcdA and TcdB Exhibit Deregulated Tight Junction Proteins.\u201d Toxins, 15(11), 643. doi:10.3390/toxins15110643 \u261e Link
Monsorno K, Ginggen K, Ivanov A, Buckinx A, Lalive AL, Tchenio A, Benson S, Vendrell M, D\u2019Alessandro A, Beule D, Pellerin L, Mameli M, Paolicelli RC (2023). \u201cLoss of microglial MCT4 leads to defective synaptic pruning and anxiety-like behavior in mice.\u201d Nature Communications, 14(1), 5749.
Zheng L, Chopra A, Weiner J, Beule D, Dommisch H, Schaefer A (2023). \u201cmiRNAs from inflamed gingiva link gene signaling to increased MET expression.\u201d Journal of Dental Research, 00220345231197984. doi:10.1177/002203452311979 \u261e Link
Mhlekude B, Postmus D, Stenzel S, Weiner J, Jansen J, Zapatero-Belinch\u00f2n FJ, Olmer R, Richter A, Heinze J, Heinemann N, M\u00fchlemann B, Schroeder S, Jones TC, M\u00fcller MA, Drosten C, Pich A, Thiel V, Martin U, Niemeyer D, Gerold G, Beule D, Goffinet C (2023). \u201cPharmacological inhibition of bromodomain and extra-terminal proteins induces an NRF-2-mediated antiviral state that is subverted by SARS-CoV-2 infection.\u201d PLoS pathogens, 19(9), e1011657. doi:10.1371/journal.ppat.1011657 \u261e Link
Benary M, Wang XD, Schmidt M, Soll D, Hilfenhaus G, Nassir M, Sigler C, Kn\u00f6dler M, Keller U, Beule D, Keilholz U, Leser U, Rieke DT (2023). \u201cLeveraging Large Language Models for Decision Support in Personalized Oncology.\u201d JAMA Network Open, 6(11), e2343689-e2343689. doi:10.1001/jamanetworkopen.2023.43689
Briest F, Noerenberg D, Hennch C, Yoshida K, Hablesreiter R, Nimo J, Sasca D, Kirchner M, Mansouri L, Inoue Y, Wiegand L, Staiger AM, Casadei B, Korkolopoulou P, Weiner J, Lopez-Guillermo A, Warth A, Schneider T, Nagy \u00c1, Klapper W, Hummel M, Kanellis G, Anagnostopoulos I, Mertins P, Bullinger L, Rosenquist R, Vassilakopoulos TP, Ott G, Ogawa S, Damm F (2023). \u201cFrequent ZNF217 mutations lead to transcriptional deregulation of interferon signal transduction via altered chromatin accessibility in B cell lymphoma.\u201d Leukemia, 1-13. doi:10.1038/s41375-023-02013-9 \u261e Link
Spranger L, Weiner J, Bredow J, Zeitz U, Grittner U, Boschmann M, Dickmann S, Stob\u00e4us N, Schwartzenberg RJ, Brachs M, Spranger J, Mai K (2023). \u201cThrifty energy phenotype predicts weight regain in postmenopausal women with overweight or obesity and is related to FGFR1 signaling.\u201d Clinical Nutrition, 42(4), 559-567. doi:10.1016/j.clnu.2023.02.020 \u261e Link
Chopra A, Song J, Weiner J, Keceli HG, Dincer PR, Cruz R, Carracedo A, Blanco J, Dommisch H, Schaefer AS (2023). \u201cRSPO4 is a potential risk gene of stages III-IV, grade C periodontitis through effects on innate immune response and oral barrier integrity.\u201d Journal of Clinical Periodontology, 50(4), 476-486. doi:10.1111/jcpe.13758 \u261e Link
Lechner L, Opitz R, Silver MJ, Krabusch PM, Prentice AM, Field MS, Stachelscheid H, Leit\u00e3o E, Schr\u00f6der C, Vallone VF, Horsthemke B, J\u00f6ckel K, Schmidt B, N\u00f6then MM, Hoffmann P, Herms S, Kleyn PW, Megges M, Blume-Peytavi U, Weiss K, Mai K, Blankenstein O, Obermayer B, Wiegand S, K\u00fchnen PKP (2023). \u201cEarly-set POMC methylation variability is accompanied by increased risk for obesity and is addressable by MC4R agonist treatment.\u201d Science Translational Medicine, 15(705), eadg1659. doi:10.1126/scitranslmed.adg1659 \u261e Link
Obermayer B, Keilholz L, Conrad T, Frentsch M, Blau I, Vuong L, Lesch S, Movasshagi K, Tietze-Stolley C, Loyal L, Henze L, Penack O, Stervbo U, Babel N, Haas S, Beule D, Bullinger L, Wittenbecher F, Na I (2023). \u201cSingle-cell clonal tracking of persistent T-cells in allogeneic hematopoietic stem cell transplantation.\u201d Frontiers in Immunology, 14. ISSN 1664-3224, doi:10.3389/fimmu.2023.1114368 \u261e Link
Gressier E, Schulte-Schrepping J, Petrov L, Brumhard S, Stubbemann P, Hiller A, Obermayer B, Spitzer J, Kostevc T, Whitney PG, Bachem A, Odainic A, van de Sandt C, Nguyen THO, Ashhurst T, Wilson K, Oates CVL, Gearing LJ, Meischel T, Hochheiser K, Greyer M, Clarke M, Kreutzenbeck M, Gabriel SS, Kastenm\u00fcller W, Kurts C, Londrigan SL, Kallies A, Kedzierska K, Hertzog PJ, Latz E, Chen YE, Radford KJ, Chopin M, Schroeder J, Kurth F, Gebhardt T, Sander LE, Sawitzki B, Schultze JL, Schmidt SV, Bedoui S (2023). \u201cCD4+ T cell calibration of antigen-presenting cells optimizes antiviral CD8+ T cell immunity.\u201d Nature Immunology, 24(6), 979-990. ISSN 1529-2916, doi:10.1038/s41590-023-01517-x \u261e Link
Loedige I, Baranovskii A, Mendonsa S, Dantsuji S, Popitsch N, Breimann L, Zerna N, Cherepanov V, Milek M, Ameres S, Chekulaeva M (2023). \u201cmRNA stability and m6A are major determinants of subcellular mRNA localization in neurons.\u201d Molecular Cell, 83(15), 2709-2725.e10. ISSN 1097-2765, doi:10.1016/j.molcel.2023.06.021 \u261e Link
Boeddrich A, Haenig C, Neuendorf N, Blanc E, Ivanov A, Kirchner M, Schleumann P, Bayraktaroglu I, Richter M, Molenda CM, Sporbert A, Zenkner M, Schnoegl S, Suenkel C, Schneider L, Rybak-Wolf A, Kochnowsky B, Byrne LM, Wild EJ, Nielsen JE, Dittmar G, Peters O, Beule D, Wanker EE (2023). \u201cA proteomics analysis of 5xFAD mouse brain regions reveals the lysosome-associated protein Arl8b as a candidate biomarker for Alzheimer's disease.\u201d Genome Med, 15, 50. doi:10.1186/s13073-023-01206-2 \u261e Link
de Bortoli T, Benary M, Horak P, Lamping M, Stintzing S, Tinhofer I, Leyvraz S, Sch\u00e4fer R, Klauschen F, Keller U, Rieke DT, Jelas I (2023). \u201cTumour mutational burden and survival with molecularly matched therapy.\u201d European Journal of Cancer, 190, 112925. doi:10.1016/j.ejca.2023.05.013 \u261e Link
Stellmach C, Sass J, Auber B, Boeker M, Wienker T, Heidel AJ, Benary M, Schumacher S, Ossowski S, Klauschen F, M\u00f6ller Y, Schmutzler R, Ustjanzew A, Werner P, Tomczak A, H\u00f6lter T, Thun S (2023). \u201cCreation of a structured molecular genomics report for Germany as a local adaption of HL7\u2019s Genomic Reporting Implementation Guide.\u201d Journal of the American Medical Informatics Association. ISSN 1527-974X, doi:10.1093/jamia/ocad061 \u261e Link
Winkler W, D\u00edaz CF, Blanc E, Napieczynska H, Langner P, Werner M, Walter B, Wollert-Wulf B, Yasuda T, Heuser A, Beule D, Mathas S, Anagnostopoulos I, Rosenwald A, Rajewsky K, Janz M (2023). \u201cMouse models of human multiple myeloma subgroups.\u201d Proceedings of the National Academy of Sciences of the United States of America, 120, e2219439120. ISSN 1091-6490, doi:10.1073/pnas.2219439120 \u261e Link
Spranger L, Weiner J, Bredow J, Zeitz U, Grittner U, Boschmann M, Dickmann S, Stobaus N, Schwartzenberg JJ, Brachs M, Spranger J, Mai K (2023). \u201cThrifty energy phenotype predicts weight regain in postmenopausal women with overweight or obesity and is related to FGFR1 signaling.\u201d Clinical Nutrition, 42, 559-567. doi:10.1016/j.clnu.2023.02.020 \u261e Link
Freitag K, Eede P, Ivanov A, Sterczyk N, Schneeberger S, Borodina T, Sauer S, Beule D, Heppner F (2023). \u201cDiverse but unique astrocytic phenotypes during embryonic stem cell differentiation, culturing and development.\u201d Commun Biol, 6, 40. doi:10.1038/s42003-023-04410-3 \u261e Link
Nieminen M, Stolpe O, Kuhring M, Weiner J, Pett P, Beule D, Holtgrewe M (2023). \u201cSODAR: managing multiomics study data and metadata.\u201d Gigascience, 12, giad052. doi:10.1093/gigascience/giad052 \u261e Link
Graff P, Wilzopolski J, Voss A, Blimkie TM, Weiner J3, Kershaw O, Panwar P, Hackett T, Br\u00f6mme D, Loyal L, Thiel A, Beule D, Hancock RE, Gruber AD, B\u00e4umer W, Hedtrich S (2023). \u201cExtracellular matrix remodeling in atopic dermatitis harnesses the onset of an asthmatic phenotype and is a potential contributor to the atopic march.\u201d Journal of Investigative Dermatology. doi:10.1016/j.jid.2023.09.278
M\u00fcller R, Freitag-Wolf S, Weiner J3, Chopra A, Top T, Dommisch H, Schaefer A (2023). \u201cCase-only design identifies interactions of genetic risk variants at SIGLEC5 and PLG with the lncRNA CTD-2353F22.1 implying the importance of periodontal wound healing for disease aetiology.\u201d J Clin Periodontol, 50, 90-101. doi:10.1111/jcpe.13712 \u261e Link
Mothes R, Pascual-Reguant A, Koehler R, Liebeskind J, Liebheit A, Bauherr S, Philipsen L, Dittmayer C, Laue M, von Manitius R, Elezkurtaj S, Durek P, Heinrich F, Heinz GA, Guerra GM, Obermayer B, Meinhardt J, Ihlow J, Radke J, Heppner FL, Enghard P, Stockmann H, Aschman T, Schneider J, Corman VM, Sander LE, Mashreghi M, Conrad T, Hocke AC, Niesner RA, Radbruch H, Hauser AE (2023). \u201cDistinct tissue niches direct lung immunopathology via CCL18 and CCL21 in severe COVID-19.\u201d Nature Communications, 14(1), 791. ISSN 2041-1723, doi:10.1038/s41467-023-36333-2 \u261e Link
Rieke DT, Schr\u00f6der S, Schafhausen P, Blanc E, Zuljan E, von der Emde B, Beule D, Keller U, Keilholz U, Klinghammer K (2023). \u201cTargeted treatment in a case series of AR+, HRAS/PIK3CA co-mutation salivary duct carcinoma.\u201d Frontiers in Oncology, 13. ISSN 2234-943X, doi:10.3389/fonc.2023.1107134
"},{"location":"misc/publication-list/#2022","title":"2022","text":"Holle J, Bartolomaeus H, L\u00f6ber U, Behrens F, Bartolomaeus TUP, Anandakumar H, Wimmer MI, Vu DL, Kuhring M, Br\u00fcning U, Maifeld A, Geisberger S, Kempa S, Schumacher F, Kleuser B, Bufler P, Querfeld U, Kitschke S, Engler D, Kuhrt LD, Drechsel O, Eckardt K, Forslund SK, Th\u00fcrmer A, McParland V, Kirwan JA, Wilck N, M\u00fcller D (2022). \u201cInflammation in Children with CKD Linked to Gut Dysbiosis and Metabolite Imbalance.\u201d Journal of the American Society of Nephrology, 33(12), 2259. ISSN 1046-6673, doi:10.1681/ASN.2022030378 \u261e Link
Baumgardt M, H\u00fclsemann M, L\u00f6wa A, Fatykhova D, Hoffmann K, Kessler M, Mieth M, Hellwig K, Frey D, Langenhagen A, Voss A, Obermayer B, Wyler E, D\u00f6kel S, Gruber AD, T\u00f6lch U, Hippenstiel S, Hocke AC, H\u00f6nzke K (2022). \u201cState-of-the-art analytical methods of viral infections in human lung organoids.\u201d PLOS ONE, 17(12), 1-8. doi:10.1371/journal.pone.0276115 \u261e Link
Pennitz P, Kirsten H, Friedrich VD, Wyler E, Goekeri C, Obermayer B, Heinz GA, Mashreghi M, B\"uttner M, Trimpert J, Landthaler M, Suttorp N, Hocke AC, Hippenstiel S, T\"onnies M, Scholz M, Kuebler WM, Witzenrath M, Hoenzke K, Nouailles G (2022). \u201cA pulmonologist's guide to perform and analyse cross-species single lung cell transcriptomics.\u201d European Respiratory Review, 31(165). ISSN 0905-9180, doi:10.1183/16000617.0056-2022 \u261e Link
Bertrams W, H\u00f6nzke K, Obermayer B, T\u00f6nnies M, Bauer TT, Schneider P, Neudecker J, R\u00fcckert JC, Stiewe T, Nist A, Eggeling S, Suttorp N, Wolff T, Hippenstiel S, Schmeck B, Hocke AC (2022). \u201cTranscriptomic comparison of primary human lung cells with lung tissue samples and the human A549 lung cell line highlights cell type specific responses during infections with influenza A virus.\u201d Scientific Reports, 12(1), 20608. ISSN 2045-2322, doi:10.1038/s41598-022-24792-4 \u261e Link
Jarick K, Topczewska P, Jakob M, Yano H, Arifuzzaman M, Gao X, Boulekou S, Stokic-Trtica V, Lecl\u00e8re P, Preu\u00dfer A, Rompe Z, Stamm A, Tsou A, Chu C, Heinrich F, Guerra G, Durek P, Ivanov A, Beule D, Helfrich S, Duerr C, K\u00fchl A, Stehle C, Romagnani C, Mashreghi M, Diefenbach A, Artis D, Klose C (2022). \u201cNon-redundant functions of group 2 innate lymphoid cells.\u201d Nature, 611(7937), 794-800. doi:10.1038/s41586-022-05395-5 \u261e Link
Ivanov A, Mattei D, Radscheit K, Compagnion A, Pett J, Herzel H, Paolicelli R, Piwecka M, Meyer U, Beule D (2022). \u201cAnalyses of circRNA Expression throughout the Light-Dark Cycle Reveal a Strong Regulation of Cdr1as, Associated with Light Entrainment in the SCN.\u201d Int J Mol Sci, 23. doi:10.3390/ijms232012347 \u261e Link
Wirges A, Bunse M, Joedicke J, Blanc E, Gudipati V, Moles M, Shiku H, Beule D, Huppa J, H\u00f6pken U, Rehm A (2022). \u201cEBAG9 silencing exerts an immune checkpoint function without aggravating adverse effects.\u201d Mol Ther, 30, 3358-3378. doi:10.1016/j.ymthe.2022.07.009 \u261e Link
H\u00f6nzke K, Obermayer B, Mache C, Fathykova D, Kessler M, D\u00f6kel S, Wyler E, Baumgardt M, L\u00f6wa A, Hoffmann K, Graff P, Schulze J, Mieth M, Hellwig K, Demir Z, Biere B, Brunotte L, Mecate-Zambrano A, Bushe J, Dohmen M, Hinze C, Elezkurtaj S, T\u00f6nnies M, Bauer T, Eggeling S, Tran H, Schneider P, Neudecker J, R\u00fcckert J, Schmidt-Ott K, Busch J, Klauschen F, Horst D, Radbruch H, Radke J, Heppner F, Corman V, Niemeyer D, M\u00fcller M, Goffinet C, Mothes R, Pascual-Reguant A, Hauser A, Beule D, Landthaler M, Ludwig S, Suttorp N, Witzenrath M, Gruber A, Drosten C, Sander L, Wolff T, Hippenstiel S, Hocke A (2022). \u201cHuman lungs show limited permissiveness for SARS-CoV-2 due to scarce ACE2 levels but virus-induced expansion of inflammatory macrophages.\u201d Eur Respir J, 60. doi:10.1183/13993003.02725-2021 \u261e Link
Freitag K, Sterczyk N, Wendlinger S, Obermayer B, Schulz J, Farztdinov V, M\u00fclleder M, Ralser M, Houtman J, Fleck L, Braeuning C, Sansevrino R, Hoffmann C, Milovanovic D, Sigrist S, Conrad T, Beule D, Heppner F, Jendrach M (2022). \u201cSpermidine reduces neuroinflammation and soluble amyloid beta in an Alzheimer's disease mouse model.\u201d J Neuroinflammation, 19, 172. doi:10.1186/s12974-022-02534-7 \u261e Link
Schalbetter S, von Arx A, Cruz-Ochoa N, Dawson K, Ivanov A, Mueller F, Lin H, Amport R, Mildenberger W, Mattei D, Beule D, F\u00f6ldy C, Greter M, Notter T, Meyer U (2022). \u201cAdolescence is a sensitive period for prefrontal microglia to act on cognitive development.\u201d Sci Adv, 8, eabi6672. doi:10.1126/sciadv.abi6672 \u261e Link
Grunert C, Willimsky G, Peuker C, Rhein S, Hansmann L, Blankenstein T, Blanc E, Beule D, Keller U, Pezzutto A, Busse A (2022). \u201cIsolation of Neoantigen-Specific Human T Cell Receptors from Different Human and Murine Repertoires.\u201d Cancers (Basel), 14. doi:10.3390/cancers14071842 \u261e Link
Dannebaum R, Suwalski P, Asgharian H, Du ZG, Lin H, Weiner J, Holtgrewe M, Thibeault C, M\u00fcller M, Wang X, Karadeniz Z, Saccomanno J, Doehn J, H\u00fcbner R, Hinzmann B, Bl\u00fcher A, Siemann S, Telman D, Suttorp N, Witzenrath M, Hippenstiel S, Skurk C, Poller W, Sander L, Beule D, Kurth F, Guettouche T, Landmesser U, Berka J, Luong K, Pa-COVID SG, Rubelt F, Heidecker B (2022). \u201cHighly multiplexed immune repertoire sequencing links multiple lymphocyte classes with severity of response to COVID-19.\u201d EClinicalMedicine, 48, 101438. doi:10.1016/j.eclinm.2022.101438 \u261e Link
Horak P, Griffith M, Danos A, Pitel B, Madhavan S, Liu X, Chow C, Williams H, Carmody L, Barrow-Laing L, Rieke D, Kreutzfeldt S, Stenzinger A, Tamborero D, Benary M, Rajagopal P, Ida C, Lesmana H, Satgunaseelan L, Merker J, Tolstorukov M, Campregher P, Warner J, Rao S, Natesan M, Shen H, Venstrom J, Roy S, Tao K, Kanagal-Shamanna R, Xu X, Ritter D, Pagel K, Krysiak K, Dubuc A, Akkari Y, Li X, Lee J, King I, Raca G, Wagner A, Li M, Plon S, Kulkarni S, Griffith O, Chakravarty D, Sonkin D (2022). \u201cStandards for the classification of pathogenicity of somatic variants in cancer (oncogenicity): Joint recommendations of Clinical Genome Resource (ClinGen), Cancer Genomics Consortium (CGC), and Variant Interpretation for Cancer Consortium (VICC).\u201d Genet Med, 24, 986-998. doi:10.1016/j.gim.2022.01.001 \u261e Link
Weiner J, Obermayer B, Beule D (2022). \u201cVenn Diagrams May Indicate Erroneous Statistical Reasoning in Transcriptomics.\u201d Front Genet, 13, 818683. doi:10.3389/fgene.2022.818683 \u261e Link
Gloaguen Y, Kirwan J, Beule D (2022). \u201cDeep Learning-Assisted Peak Curation for Large-Scale LC-MS Metabolomics.\u201d Anal Chem, 94, 4930-4937. doi:10.1021/acs.analchem.1c02220 \u261e Link
Chopra A, Mueller R, Weiner J3, Rosowski J, Dommisch H, Grohmann E, Schaefer A (2022). \u201cBACH1 Binding Links the Genetic Risk for Severe Periodontitis with ST8SIA1.\u201d J Dent Res, 101, 93-101. doi:10.1177/00220345211017510 \u261e Link
Bischoff P, Trinks A, Wiederspahn J, Obermayer B, Pett J, Jurmeister P, Elsner A, Dziodzio T, R\u00fcckert J, Neudecker J, Falk C, Beule D, Sers C, Morkel M, Horst D, Klauschen F, Bl\u00fcthgen N (2022). \u201cThe single-cell transcriptional landscape of lung carcinoid tumors.\u201d Int J Cancer, 150, 2058-2071. doi:10.1002/ijc.33995 \u261e Link
Georg P, Astaburuaga-Garc\u00eda R, Bonaguro L, Brumhard S, Michalick L, Lippert L, Kostevc T, G\u00e4bel C, Schneider M, Streitz M, Demichev V, Gem\u00fcnd I, Barone M, Tober-Lau P, Helbig E, Hillus D, Petrov L, Stein J, Dey H, Paclik D, Iwert C, M\u00fclleder M, Aulakh S, Djudjaj S, B\u00fclow R, Mei H, Schulz A, Thiel A, Hippenstiel S, Saliba A, Eils R, Lehmann I, Mall M, Stricker S, R\u00f6hmel J, Corman V, Beule D, Wyler E, Landthaler M, Obermayer B, von Stillfried S, Boor P, Demir M, Wesselmann H, Suttorp N, Uhrig A, M\u00fcller-Redetzky H, Nattermann J, Kuebler W, Meisel C, Ralser M, Schultze J, Aschenbrenner A, Thibeault C, Kurth F, Sander L, Bl\u00fcthgen N, Sawitzki B, PA-COVID-19 SG (2022). \u201cComplement activation induces excessive T cell cytotoxicity in severe COVID-19.\u201d Cell, 185(3), 493-512.e25. ISSN 0092-8674, doi:10.1016/j.cell.2021.12.040 \u261e Link
Zinnall U, Milek M, Minia I, Vieira-Vieira C, M\u00fcller S, Mastrobuoni G, Hazapis O, Del GS, Schwefel D, Bley N, Voigt F, Chao J, Kempa S, H\u00fcttelmaier S, Selbach M, Landthaler M (2022). \u201cHDLBP binds ER-targeted mRNAs by multivalent interactions to promote protein synthesis of transmembrane and secreted proteins.\u201d Nat Commun, 13(1), 2727. doi:10.1038/s41467-022-30322-7 \u261e Link
D\u00f6rr D, Obermayer B, Weiner J, Zimmermann K, Anania C, Wagner L, Lyras E, Sapozhnikova V, Lara-Astiaso D, Pr\u00f3sper F, Lang R, Lupi\u00e1\u00f1ez D, Beule D, H\u00f6pken U, Leutz A, Mildner A (2022). \u201cC/EBP\u03b2 regulates lipid metabolism and Pparg isoform 2 expression in alveolar macrophages.\u201d Sci Immunol, 7(75), eabj0140. doi:10.1126/sciimmunol.abj0140 \u261e Link
Kedziora S, Obermayer B, Sugulle M, Herse F, Kr\u00e4ker K, Haase N, Langmia I, M\u00fcller D, Staff A, Beule D, Dechend R (2022). \u201cPlacental Transcriptome Profiling in Subtypes of Diabetic Pregnancies Is Strongly Confounded by Fetal Sex.\u201d Int J Mol Sci, 23(23). ISSN 1422-0067, doi:10.3390/ijms232315388 \u261e Link
Pennitz P, Kirsten H, Wyler E, Goekeri C, Obermayer B, Heinz GA, Mashreghi MF, Trimpert J, Landthaler M, Suttorp N, Hocke AC, Hippenstiel S, T\u00f6nnies M, Scholz M, Kuebler WM, Witzenrath M, Hoenzke K, Nouailles G (2022). \u201cIntegrational approaches for cross-species analysis of lung pathologies at single-cell resolution.\u201d Pneumologie, 76(S 01), FV 285. FV 285.
Hoffmann K, Obermayer B, H\u00f6nzke K, Fatykhova D, Demir Z, L\u00f6wa A, Alves L, Wyler E, Lopez-Rodriguez E, Mieth M, Baumgardt M, Hoppe J, Firsching T, T\u00f6nnies M, Bauer T, Eggeling S, Tran H, Schneider P, Neudecker J, R\u00fcckert J, Gruber A, Ochs M, Landthaler M, Beule D, Suttorp N, Hippenstiel S, Hocke A, Kessler M (2022). \u201cHuman alveolar progenitors generate dual lineage bronchioalveolar organoids.\u201d Commun Biol, 5(1), 875. ISSN 2399-3642, doi:10.1038/s42003-022-03828-5 \u261e Link
Neitzel H, Varon R, Chughtai S, Dartsch J, Dutrannoy-T\u00f6nsing V, N\u00fcrnberg P, N\u00fcrnberg G, Schweiger M, Digweed M, Hildebrand G, Hackmann K, Holtgrewe M, Sarioglu N, Schulze B, Horn D, Sperling K (2022). \u201cTransmission ratio distortion of mutations in the master regulator of centriole biogenesis PLK4.\u201d Hum Genet, 141(11), 1785-1794. doi:10.1007/s00439-022-02461-w \u261e Link
Foddis M, Blumenau S, Holtgrewe M, Paquette K, Westra K, Alonso I, Macario M, Morgadinho A, Velon A, Santo G, Santana I, M\u00f6nk\u00e4re S, Kuuluvainen L, Schleutker J, P\u00f6yh\u00f6nen M, Myllykangas L, Pavlovic A, Kostic V, Dobricic V, Lohmann E, Hanagasi H, Santos M, Guven G, Bilgic B, Bras J, Beule D, Dirnagl U, Guerreiro R, Sassi C (2022). \u201cTREX1 p.A129fs and p.Y305C variants in a large multi-ethnic cohort of CADASIL-like unrelated patients.\u201d doi:10.1016/j.neurobiolaging.2022.11.013 \u261e Link
Boschann F, Moreno D, Mensah M, Sczakiel H, Skipalova K, Holtgrewe M, Mundlos S, Fischer-Zirnsak B (2022). \u201cXq27.1 palindrome mediated interchromosomal insertion likely causes familial congenital bilateral laryngeal abductor paralysis (Plott syndrome).\u201d J Hum Genet, 67(7), 405-410. doi:10.1038/s10038-022-01018-z \u261e Link
Seidel F, Laser K, Klingel K, Dartsch J, Theisen S, Pickardt T, Holtgrewe M, G\u00e4rtner A, Berger F, Beule D, Milting H, Schubert S, Klaassen S, K\u00fchnisch J (2022). \u201cPathogenic Variants in Cardiomyopathy Disorder Genes Underlie Pediatric Myocarditis-Further Impact of Heterozygous Immune Disorder Gene Variants?\u201d J Cardiovasc Dev Dis, 9(7), 216. doi:10.3390/jcdd9070216 \u261e Link
May V, Koch L, Fischer-Zirnsak B, Horn D, Gehle P, Kornak U, Beule D, Holtgrewe M (2022). \u201cClearCNV: CNV calling from NGS panel data in the presence of ambiguity and noise.\u201d Bioinformatics, 38(16), 3871-3876. doi:10.1093/bioinformatics/btac418 \u261e Link
Sch\u00f6pflin R, Melo U, Moeinzadeh H, Heller D, Laupert V, Hertzberg J, Holtgrewe M, Alavi N, Klever M, Jungnitsch J, Comak E, T\u00fcrkmen S, Horn D, Duffourd Y, Faivre L, Callier P, Sanlaville D, Zuffardi O, Tenconi R, Kurtas N, Giglio S, Prager B, Latos-Bielenska A, Vogel I, Bugge M, Tommerup N, Spielmann M, Vitobello A, Kalscheuer V, Vingron M, Mundlos S (2022). \u201cIntegration of Hi-C with short and long-read genome sequencing reveals the structure of germline rearranged genomes.\u201d Nat Commun, 13(1), 6470. doi:10.1038/s41467-022-34053-7 \u261e Link
Rieke D, de BT, Horak P, Lamping M, Benary M, Jelas I, R\u00fcter G, Berger J, Zettwitz M, Kagelmann N, Kind A, Fabian F, Beule D, Glimm H, Brors B, Stenzinger A, Fr\u00f6hling S, Keilholz U (2022). \u201cFeasibility and outcome of reproducible clinical interpretation of high-dimensional molecular data: a comparison of two molecular tumor boards.\u201d BMC Med, 20, 367. ISSN 17417015, doi:10.1186/s12916-022-02560-5 \u261e Link
Hinze C, Kocks C, Leiz J, Karaiskos N, Boltengagen A, Cao S, Skopnik C, Klocke J, Hardenberg J, Stockmann H, Gotthardt I, Obermayer B, Haghverdi L, Wyler E, Landthaler M, Bachmann S, Hocke A, Corman V, Busch J, Schneider W, Himmerkus N, Bleich M, Eckardt K, Enghard P, Rajewsky N, Schmidt-Ott K (2022). \u201cSingle-cell transcriptomics reveals common epithelial response patterns in human acute kidney injury.\u201d Genome Med, 14, 103. doi:10.1186/s13073-022-01108-9 \u261e Link
Richter G, Wagner G, Reichenmiller K, Staufenbiel I, Martins O, L\u00f6scher B, Holtgrewe M, Jepsen S, Dommisch H, Schaefer A (2022). \u201cExome Sequencing of 5 Families with Severe Early-Onset Periodontitis.\u201d J Dent Res, 101(2), 151-157. doi:10.1177/00220345211029266 \u261e Link
Coutelier M, Holtgrewe M, J\u00e4ger M, Fl\u00f6ttman R, Mensah M, Spielmann M, Krawitz P, Horn D, Beule D, Mundlos S (2022). \u201cCombining callers improves the detection of copy number variants from whole-genome sequencing.\u201d Eur J Hum Genet, 30(2), 178-186. doi:10.1038/s41431-021-00983-x \u261e Link
Vogt G, Verheyen S, Schwartzmann S, Ehmke N, Potratz C, Schwerin-Nagel A, Plecko B, Holtgrewe M, Seelow D, Blatterer J, Speicher M, Kornak U, Horn D, Mundlos S, Fischer-Zirnsak B, Boschann F (2022). \u201cBiallelic truncating variants in ATP9A cause a novel neurodevelopmental disorder involving postnatal microcephaly and failure to thrive.\u201d J Med Genet, 59(7), 662-668. doi:10.1136/jmedgenet-2021-107843 \u261e Link
Kossen T, Hirzel MA, Madai VI, Boenisch F, Hennemuth A, Hildebrand K, Pokutta S, Sharma K, Hilbert A, Sobesky J, Galinovic I, Khalil AA, Fiebach JB and Frey D. Toward Sharing Brain Images: Differentially Private TOF-MRA Images With Segmentation Labels Using Generative Adversarial Networks. Frontiers in Artificial Intelligence. 5 (2022). issn: 2624-8212. doi: 10.3389/frai.2022.813842
"},{"location":"misc/publication-list/#2021","title":"2021","text":"Schinke C, Fernandez VV, Ivanov A, Peng Y, K\u00f6rtvelyessy P, Nolte L, Huehnchen P, Beule D, Stachelscheid H, Boehmerle W, Endres M (2021). \u201cDataset for: Modeling chemotherapy induced neurotoxicity with human induced pluripotent stem cell (iPSC)-derived sensory neurons.\u201d Data Brief, 38, 107320. doi:10.1016/j.dib.2021.107320 \u261e Link
Weiner J, Suwalski P, Holtgrewe M, Rakitko A, Thibeault C, M\u00fcller M, Patriki D, Quedenau C, Kr\u00fcger U, Ilinsky V, Popov I, Balnis J, Jaitovich A, Helbig E, Lippert L, Stubbemann P, Real L, Mac\u00edas J, Pineda J, Fernandez-Fuertes M, Wang X, Karadeniz Z, Saccomanno J, Doehn J, H\u00fcbner R, Hinzmann B, Salvo M, Blueher A, Siemann S, Jurisic S, Beer J, Rutishauser J, Wiggli B, Schmid H, Danninger K, Binder R, Corman V, M\u00fchlemann B, Arjun AR, Fragiadakis G, Mick E, Comet C, Calfee C, Erle D, Hendrickson C, Kangelaris K, Krummel M, Woodruff P, Langelier C, Venkataramani U, Garc\u00eda F, Zyla J, Drosten C, Alice B, Jones T, Suttorp N, Witzenrath M, Hippenstiel S, Zemojtel T, Skurk C, Poller W, Borodina T, Pa-Covid S, Ripke S, Sander L, Beule D, Landmesser U, Guettouche T, Kurth F, Heidecker B (2021). \u201cIncreased risk of severe clinical course of COVID-19 in carriers of HLA-C\\04:01.\u201d EClinicalMedicine, 40, 101099. doi:10.1016/j.eclinm.2021.101099 \u261e Link
Domaszewska T, Zyla J, Otto R, Kaufmann S, Weiner J (2021). \u201cGene Set Enrichment Analysis Reveals Individual Variability in Host Responses in Tuberculosis Patients.\u201d Front Immunol, 12, 694680. doi:10.3389/fimmu.2021.694680 \u261e Link
Nouailles G, Wyler E, Pennitz P, Postmus D, Vladimirova D, Kazmierski J, Pott F, Dietert K, Muelleder M, Farztdinov V, Obermayer B, Wienhold S, Andreotti S, Hoefler T, Sawitzki B, Drosten C, Sander LE, Suttorp N, Ralser M, Beule D, Gruber AD, Goffinet C, Landthaler M, Trimpert J, Witzenrath M (2021). \u201cSingle-cell-sequencing in SARS-COV-2-infected hamsters sheds light on endothelial cell involvement in COVID-19.\u201d European Respiratory Journal. doi:10.1183/13993003.congress-2021.PA2355 \u261e Link
Bischoff P, Trinks A, Obermayer B, Pett J, Wiederspahn J, Uhlitz F, Liang X, Lehmann A, Jurmeister P, Elsner A, Dziodzio T, R\u00fcckert J, Neudecker J, Falk C, Beule D, Sers C, Morkel M, Horst D, Bl\u00fcthgen N, Klauschen F (2021). \u201cSingle-cell RNA sequencing reveals distinct tumor microenvironmental patterns in lung adenocarcinoma.\u201d Oncogene, 40, 6748-6758. doi:10.1038/s41388-021-02054-3 \u261e Link
Uhlitz F, Bischoff P, Peidli S, Sieber A, Trinks A, L\u00fcthen M, Obermayer B, Blanc E, Ruchiy Y, Sell T, Mamlouk S, Arsie R, Wei T, Klotz-Noack K, Schwarz R, Sawitzki B, Kamphues C, Beule D, Landthaler M, Sers C, Horst D, Bl\u00fcthgen N, Morkel M (2021). \u201cMitogen-activated protein kinase activity drives cell trajectories in colorectal cancer.\u201d EMBO Mol Med, 13, e14123. doi:10.15252/emmm.202114123 \u261e Link
Nouailles G, Wyler E, Pennitz P, Postmus D, Vladimirova D, Kazmierski J, Pott F, Dietert K, Muelleder M, Farztdinov V, Obermayer B, Wienhold S, Andreotti S, Hoefler T, Sawitzki B, Drosten C, Sander L, Suttorp N, Ralser M, Beule D, Gruber A, Goffinet C, Landthaler M, Trimpert J, Witzenrath M (2021). \u201cTemporal omics analysis in Syrian hamsters unravel cellular effector responses to moderate COVID-19.\u201d Nat Commun, 12, 4869. doi:10.1038/s41467-021-25030-7 \u261e Link
Paterson S, Kar S, Ung S, Gardener Z, Bergstrom E, Ascough S, Kalyan M, Zyla J, Maertzdorf J, Mollenkopf H, Weiner J, Jozwik A, Jarvis H, Jha A, Nicholson B, Veldman T, Woods C, Mallia P, Kon O, Kaufmann S, Openshaw P, Chiu C (2021). \u201cInnate-like Gene Expression of Lung-Resident Memory CD8(+) T Cells during Experimental Human Influenza: A Clinical Study.\u201d Am J Respir Crit Care Med, 204, 826-841. doi:10.1164/rccm.202103-0620OC \u261e Link
Rajasekaran S, Siddiqui J, Rakijas J, Nicolay B, Lin C, Khan E, Patel R, Morris R, Wyler E, Boukhali M, Balasubramanyam J, Ranjith KR, van Rechem C, Vogel C, Elchuri S, Landthaler M, Obermayer B, Haas W, Dyson N, Miles W (2021). \u201cIntegrated multi-omics analysis of RB-loss identifies widespread cellular programming and synthetic weaknesses.\u201d Commun Biol, 4, 977. doi:10.1038/s42003-021-02495-2 \u261e Link
Schinke C, Fernandez VV, Ivanov A, Peng Y, K\u00f6rtvelyessy P, Nolte L, Huehnchen P, Beule D, Stachelscheid H, Boehmerle W, Endres M (2021). \u201cModeling chemotherapy induced neurotoxicity with human induced pluripotent stem cell (iPSC) -derived sensory neurons.\u201d Neurobiol Dis, 155, 105391. doi:10.1016/j.nbd.2021.105391 \u261e Link
Pei G, Zyla J, He L, Moura-Alves P, Steinle H, Saikali P, Lozza L, Nieuwenhuizen N, Weiner J, Mollenkopf H, Ellwanger K, Arnold C, Duan M, Dagil Y, Pashenkov M, Boneca I, Kufer T, Dorhoi A, Kaufmann S (2021). \u201cCellular stress promotes NOD\u00bd-dependent inflammation via the endogenous metabolite sphingosine-1-phosphate.\u201d EMBO J, 40, e106272. doi:10.15252/embj.2020106272 \u261e Link
Herda S, Heimann A, Obermayer B, Ciraolo E, Althoff S, Ru\u00df J, Grunert C, Busse A, Bullinger L, Pezzutto A, Blankenstein T, Beule D, Na I (2021). \u201cLong-term in vitro expansion ensures increased yield of central memory T cells as perspective for manufacturing challenges.\u201d Int J Cancer, 148, 3097-3110. doi:10.1002/ijc.33523 \u261e Link
Kraushar M, Krupp F, Harnett D, Turko P, Ambrozkiewicz M, Sprink T, Imami K, G\u00fcnnigmann M, Zinnall U, Vieira-Vieira C, Schaub T, M\u00fcnster-Wandowski A, B\u00fcrger J, Borisova E, Yamamoto H, Rasin M, Ohler U, Beule D, Mielke T, Tarabykin V, Landthaler M, Kramer G, Vida I, Selbach M, Spahn C (2021). \u201cProtein Synthesis in the Developing Neocortex at Near-Atomic Resolution Reveals Ebp1-Mediated Neuronal Proteostasis at the 60S Tunnel Exit.\u201d Mol Cell, 81, 304-322.e16. doi:10.1016/j.molcel.2020.11.037 \u261e Link
Heyckendorf J, Marwitz S, Reimann M, Avsar K, DiNardo A, G\u00fcnther G, Hoelscher M, Ibraim E, Kalsdorf B, Kaufmann S, Kontsevaya I, van Leth F, Mandalakas A, Maurer F, M\u00fcller M, Nitschkowski D, Olaru I, Popa C, Rachow A, Rolling T, Rybniker J, Salzer H, Sanchez-Carballo P, Schuhmann M, Schaub D, Spinu V, Su\u00e1rez I, Terhalle E, Unnewehr M, Weiner J3, Goldmann T, Lange C (2021). \u201cPrediction of anti-tuberculosis treatment duration based on a 22-gene transcriptomic model.\u201d Eur Respir J, 58. doi:10.1183/13993003.03492-2020 \u261e Link
Bao X, Weiner J3, Meckes O, Dommisch H, Schaefer A (2021). \u201cEntamoeba gingivalis Exerts Severe Pathogenic Effects on the Oral Mucosa.\u201d J Dent Res, 100, 771-776. doi:10.1177/00220345211004498 \u261e Link
Eckold C, Kumar V, Weiner J, Alisjahbana B, Riza A, Ronacher K, Coronel J, Kerry-Barnard S, Malherbe S, Kleynhans L, Stanley K, Ruslami R, Ioana M, Ugarte-Gil C, Walzl G, van Crevel R, Wijmenga C, Critchley J, Dockrell H, Cliff J, TANDEM c (2021). \u201cImpact of Intermediate Hyperglycemia and Diabetes on Immune Dysfunction in Tuberculosis.\u201d Clin Infect Dis, 72, 69-78. doi:10.1093/cid/ciaa751 \u261e Link
Wyler E, M\u00f6sbauer K, Franke V, Diag A, Gottula L, Arsi\u00e8 R, Klironomos F, Koppstein D, H\u00f6nzke K, Ayoub S, Buccitelli C, Hoffmann K, Richter A, Legnini I, Ivanov A, Mari T, Del GS, Papies J, Praktiknjo S, Meyer T, M\u00fcller M, Niemeyer D, Hocke A, Selbach M, Akalin A, Rajewsky N, Drosten C, Landthaler M (2021). \u201cTranscriptomic profiling of SARS-CoV-2 infected human cell lines identifies HSP90 as target for COVID-19 therapy.\u201d iScience, 24, 102151. doi:10.1016/j.isci.2021.102151 \u261e Link
M\u00f6lder F, Jablonski K, Letcher B, Hall M, Tomkins-Tinch C, Sochat V, Forster J, Lee S, Twardziok S, Kanitz A, Wilm A, Holtgrewe M, Rahmann S, Nahnsen S, K\u00f6ster J (2021). \u201cSustainable data analysis with Snakemake.\u201d F1000Res, 10, 33. doi:10.12688/f1000research.29032.2 \u261e Link
Messerschmidt C, Foddis M, Blumenau S, M\u00fcller S, Bentele K, Holtgrewe M, Kun-Rodrigues C, Alonso I, do CMM, Morgadinho A, Velon A, Santo G, Santana I, M\u00f6nk\u00e4re S, Kuuluvainen L, Schleutker J, P\u00f6yh\u00f6nen M, Myllykangas L, Senatore A, Berchtold D, Winek K, Meisel A, Pavlovic A, Kostic V, Dobricic V, Lohmann E, Hanagasi H, Guven G, Bilgic B, Bras J, Guerreiro R, Beule D, Dirnagl U, Sassi C (2021). \u201cPHACTR1 genetic variability is not critical in small vessel ischemic disease patients and PcomA recruitment in C57BL/6J mice.\u201d Sci Rep, 11(1), 6072. doi:10.1038/s41598-021-84919-x \u261e Link
Zarate Y, Uehara T, Abe K, Oginuma M, Harako S, Ishitani S, Lehesjoki A, Bierhals T, Kloth K, Ehmke N, Horn D, Holtgrewe M, Anderson K, Viskochil D, Edgar-Zarate C, Sacoto M, Schnur R, Morrow M, Sanchez-Valle A, Pappas J, Rabin R, Muona M, Anttonen A, Platzer K, Luppe J, Gburek-Augustat J, Kaname T, Okamoto N, Mizuno S, Kaido Y, Ohkuma Y, Hirose Y, Ishitani T, Kosaki K (2021). \u201cCDK19-related disorder results from both loss-of-function and gain-of-function de novo missense variants.\u201d Genet Med, 23(6), 1050-1057. doi:10.1038/s41436-020-01091-9 \u261e Link
Vogt G, El CN, Herczegfalvi \u00c1, K\u00f6lbel H, Lekaj A, Schara U, Holtgrewe M, Krause S, Horvath R, Schuelke M, H\u00fcbner C, Mundlos S, Roos A, Lochm\u00fcller H, Karcagi V, Kornak U, Fischer-Zirnsak B (2021). \u201cExpanding the clinical and molecular spectrum of ATP6V1A related metabolic cutis laxa.\u201d J Inherit Metab Dis, 44(4), 972-986. doi:10.1002/jimd.12341 \u261e Link
Yildirim F, Foddis M, Blumenau S, M\u00fcller S, Kajetan B, Holtgrewe M, Kola V, Beule D, Sassi C (2021). \u201cShared and oppositely regulated transcriptomic signatures in Huntington's disease and brain ischemia confirm known and unveil novel potential neuroprotective genes.\u201d Neurobiol Aging, 104, 122.e1-122.e17. doi:10.1016/j.neurobiolaging.2021.03.001 \u261e Link
Elsner J, Mensah M, Holtgrewe M, Hertzberg J, Bigoni S, Busche A, Coutelier M, de SD, El\u00e7ioglu N, Filges I, Gerkes E, Girisha K, Graul-Neumann L, Jamsheer A, Krawitz P, Kurth I, Markus S, Megarbane A, Reis A, Reuter M, Svoboda D, Teller C, Tuysuz B, T\u00fcrkmen S, Wilson M, Woitschach R, Vater I, Caliebe A, H\u00fclsemann W, Horn D, Mundlos S, Spielmann M (2021). \u201cGenome sequencing in families with congenital limb malformations.\u201d Hum Genet, 140(8), 1229-1239. doi:10.1007/s00439-021-02295-y \u261e Link
Seidel F, Holtgrewe M, Al-Wakeel-Marquard N, Opgen-Rhein B, Dartsch J, Herbst C, Beule D, Pickardt T, Klingel K, Messroghli D, Berger F, Schubert S, K\u00fchnisch J, Klaassen S (2021). \u201cPathogenic Variants Associated With Dilated Cardiomyopathy Predict Outcome in Pediatric Myocarditis.\u201d Circ Genom Precis Med, 14(4), e003250. doi:10.1161/CIRCGEN.120.003250 \u261e Link
Melo U, Piard J, Fischer-Zirnsak B, Klever M, Sch\u00f6pflin R, Mensah M, Holtgrewe M, Arbez-Gindre F, Martin A, Guigue V, Gaillard D, Landais E, Roze V, Kremer V, Ramanah R, Cabrol C, Harms F, Kornak U, Spielmann M, Mundlos S, van Maldergem L (2021). \u201cComplete lung agenesis caused by complex genomic rearrangements with neo-TAD formation at the SHH locus.\u201d Hum Genet, 140(10), 1459-1469. doi:10.1007/s00439-021-02344-6 \u261e Link
Sczakiel H, H\u00fclsemann W, Holtgrewe M, Abad-Perez A, Elsner J, Schwartzmann S, Horn D, Spielmann M, Mundlos S, Mensah M (2021). \u201cGLI3 variants causing isolated polysyndactyly are not restricted to the protein's C-terminal third.\u201d Clin Genet, 100(6), 758-765. doi:10.1111/cge.14059 \u261e Link
Li, N., Hollunder, B., Baldermann, J. C., Kibleur, A., Treu, S., Akram, H., Al-Fatly, B., Strange, B. A., Barcia, J. A., Zrinzo, L., Joyce, E. M., Chabardes, S., Visser-Vandewalle, V., Polosan, M., Kuhn, J., K\u00fchn, A. A., & Horn, A. (2021). A Unified Functional Network Target for Deep Brain Stimulation in Obsessive-Compulsive Disorder. Biological Psychiatry. doi: 10.1016/j.biopsych.2021.04.006
Bressem KK, Vahldiek JL, Adams L, Niehues SM, Haibel H, Rodriguez VR, Torgutalp M, Protopopov M, Proft F, Rademacher J, Sieper J, Rudwaleit M, Hamm B, Makowski MR, Hermann KG, Poddubnyy D. Deep learning for detection of radiographic sacroiliitis: achieving expert-level performance. Arthritis Res Ther. 2021 Apr 8;23(1):106. doi: 10.1186/s13075-021-02484-0
Kossen T, Subramaniam P, Madai VI, Hennemuth A, Hildebrand K, Hilbert A, Sobesky J, Livne M, Galinovic I, Khalil AA, Fiebach JB, Frey D. Synthesizing anonymized and labeled TOF-MRA patches for brain vessel segmentation using generative adversarial networks. Computers in Biology and Medicine. 2021 Apr 131,104254. doi: 10.1016/j.compbiomed.2021.104254
Paraskevopoulou S., K\u00e4fer S., Zirkel F., Donath A., Petersen M., Liu S., Zhou X., Drosten C., Misof B., Junglen S. (2021). \"Viromics of extant insect orders unveil the evolution of the flavi-like superfamily.\" Virus Evolution 2021 Mar 30. doi: 10.1093/ve/veab030
Thomas Krannich, W Timothy J White, Sebastian Niehus, Guillaume Holley, Bjarni V Halld\u00f3rsson, Birte Kehr, Population-scale detection of non-reference sequence variants using colored de Bruijn graphs, Bioinformatics, 2021, btab749, doi: 10.1093/bioinformatics/btab749
Julia Markowski, Rieke Kempfer, Alexander Kukalev, Ibai Irastorza-Azcarate, Gesa Loof, Birte Kehr, Ana Pombo, Sven Rahmann, Roland F Schwarz, GAMIBHEAR: whole-genome haplotype reconstruction from Genome Architecture Mapping data, Bioinformatics, Volume 37, Issue 19, 1 October 2021, Pages 3128\u20133135. doi: 10.1093/bioinformatics/btab238
"},{"location":"misc/publication-list/#2020","title":"2020","text":"Kr\u00fctzfeldt LM, Schubach M, Kircher M. The impact of different negative training data on regulatory sequence predictions. PLoS One. 2020 Dec 1;15(12):e0237412. doi: 10.1371/journal.pone.0237412.
Klotz-Noack K, Klinger B, Rivera M, Bublitz N, Uhlitz F, Riemer P, L\u00fcthen M, Sell T, Kasack K, Gastl B, Ispasanie SSS, Simon T, Janssen N, Schwab M, Zuber J, Horst D, Bl\u00fcthgen N, Sch\u00e4fer R, Morkel M, Sers C. SFPQ Depletion Is Synthetically Lethal with BRAFV600E in Colorectal Cancer Cells. Cell Rep. 2020 Sep 22;32(12):108184. doi: 10.1016/j.celrep.2020.108184.
Kleinert, P., Martin, B., & Kircher, M. (2020). \"HemoMIPs\u2014Automated analysis and result reporting pipeline for targeted sequencing data.\" PLOS Computational Biology, 16(6), e1007956. doi: 10.1371/journal.pcbi.1007956
Ehmke, N.; Cusmano-Ozog, K.; Koenig, R.; Holtgrewe, M.; Nur, B.; Mihci, E.; Babcock, H.; Gonzaga-Jauregui, C.; Overton, J. D.; Xiao, J.; et al. Biallelic Variants in KYNU Cause a Multisystemic Syndrome with Hand Hyperphalangism. Bone 2020, 115219. doi: 10.1016/j.bone.2019.115219.
Niehus, S.; J\u00f3nsson, H.; Sch\u00f6nberger, J.; Bj\u00f6rnsson, E.; Beyter, D.; Eggertsson, H.P.; Sulem, P.; Stef\u00e1nsson, K.; Halld\u00f3rsson, B.V.; Kehr, B. PopDel identifies medium-size deletions jointly in tens of thousands of genomes. bioRxiv 2020, 10.1101/740225 doi: 10.1101/740225
Gordon, M. G., Inoue, F., Martin, B., Schubach, M., Agarwal, V., Whalen, S., ... & Kreimer, A. (2020). \"lentiMPRA and MPRAflow for high-throughput functional characterization of gene regulatory elements.\" Nature Protocols, 15(8), 2387-2412. doi: 10.1038/s41596-020-0333-5
Paraskevopoulou S., Pirzer F., Goldmann N., Schmid J., Corman V.M., Gottula L.T.,Schroeder S., Rasche A., Muth D., Drexler J.F., Heni A.C., Eibner G.J., Page R.A., Jones T.C., M\u00fcllerM.A., Sommer S., Glebe D., and Drosten C. (2020). \"Mammalian deltavirus without hepadnavirus coinfection in the neotropical rodent Proechimys semispinosus.\" Proceedings of the National Academy of Sciences 2020 Jul 28;117(30):17977-17983. doi: 10.1073/pnas.2006750117.
Mattei D, Ivanov A, van Oostrum M, Pantelyushin S, Richetto J, Mueller F, Beffinger M, Schellhammer L, Vom BJ, Wollscheid B, Beule D, Paolicelli R, Meyer U (2020). \u201cEnzymatic Dissociation Induces Transcriptional and Proteotype Bias in Brain Cell Populations.\u201d Int J Mol Sci, 21. doi:10.3390/ijms21217944 \u261e Link
Kuhring M, Eisenberger A, Schmidt V, Kr\u00e4nkel N, Leistner D, Kirwan J, Beule D (2020). \u201cConcepts and Software Package for Efficient Quality Control in Targeted Metabolomics Studies: MeTaQuaC.\u201d Anal Chem, 92, 10241-10245. doi:10.1021/acs.analchem.0c00136 \u261e Link
Koche R, Rodriguez-Fos E, Helmsauer K, Burkert M, MacArthur I, Maag J, Chamorro R, Munoz-Perez N, Puiggr\u00f2s M, Dorado GH, Bei Y, R\u00f6efzaad C, Bardinet V, Szymansky A, Winkler A, Thole T, Timme N, Kasack K, Fuchs S, Klironomos F, Thiessen N, Blanc E, Schmelz K, K\u00fcnkele A, Hundsd\u00f6rfer P, Rosswog C, Theissen J, Beule D, Deubzer H, Sauer S, Toedling J, Fischer M, Hertwig F, Schwarz R, Eggert A, Torrents D, Schulte J, Henssen A (2020). \u201cExtrachromosomal circular DNA drives oncogenic genome remodeling in neuroblastoma.\u201d Nat Genet, 52, 29-34. doi:10.1038/s41588-019-0547-z \u261e Link
Holtgrewe M, Stolpe O, Nieminen M, Mundlos S, Knaus A, Kornak U, Seelow D, Segebrecht L, Spielmann M, Fischer-Zirnsak B, Boschann F, Scholl U, Ehmke N, Beule D (2020). \u201cVarFish: comprehensive DNA variant analysis for diagnostics and research.\u201d Nucleic Acids Res, 48, W162-W169. doi:10.1093/nar/gkaa241 \u261e Link
Obermayer B, Holtgrewe M, Nieminen M, Messerschmidt C, Beule D (2020). \u201cSCelVis: exploratory single cell data analysis on the desktop and in the cloud.\u201d PeerJ, 8, e8607. doi:10.7717/peerj.8607 \u261e Link
Nieminen M, Stolpe O, Schumann F (2020). \u201cSODAR Core: a Django-based framework for scientific data management and analysis web apps.\u201d Journal of Open Source Software, 5(55), 1584. doi:10.21105/joss.01584 \u261e Link
Akpa C, Kleo K, Oker E, Tomaszewski N, Messerschmidt C, L\u00f3pez C, Wagener R, Oehl-Huber K, Dettmer K, Schoeler A, Lenze D, Oefner P, Beule D, Siebert R, Capper D, Dimitrova L, Hummel M (2020). \u201cAcquired resistance to DZNep-mediated apoptosis is associated with copy number gains of AHCY in a B-cell lymphoma model.\u201d BMC Cancer, 20, 427. doi:10.1186/s12885-020-06937-8 \u261e Link
Messerschmidt C, Obermayer B, Klinghammer K, Ochsenreither S, Treue D, Stenzinger A, Glimm H, Fr\u00f6hling S, Kindler T, Brandts C, Schulze-Osthoff K, Weichert W, Tinhofer I, Klauschen F, Keilholz U, Beule D, Rieke D (2020). \u201cDistinct immune evasion in APOBEC-enriched, HPV-negative HNSCC.\u201d Int J Cancer, 147, 2293-2302. doi:10.1002/ijc.33123 \u261e Link
Boschann F, Fischer-Zirnsak B, Wienker T, Holtgrewe M, Seelow D, Eichhorn B, D\u00f6hnert S, Fahsold R, Horn D, Graul-Neumann L (2020). \u201cAn intronic splice site alteration in combination with a large deletion affecting VPS13B (COH1) causes Cohen syndrome.\u201d Eur J Med Genet, 63, 103973. doi:10.1016/j.ejmg.2020.103973 \u261e Link
Rieke D, Klinghammer K, Obermayer B, Blanc E, Messerschmidt C, J\u00f6hrens K, Klauschen F, Tinhofer I, Beule D, Keilholz U, Ochsenreither S (2020). \u201cAssociation of a STK11/KEAP1-Mutation Gene Expression Signature in Lung Adenocarcinoma with Immune Desertion in Squamous Cell Carcinomas and Mediation by NFE2L2 Deregulation.\u201d J Clin Oncol, 38(15_suppl), 3082-3082. doi:10.1200/JCO.2020.38.15_suppl.3082 \u261e Link
Ehmke N, Cusmano-Ozog K, Koenig R, Holtgrewe M, Nur B, Mihci E, Babcock H, Gonzaga-Jauregui C, Overton J, Xiao J, Martinez A, Muenke M, Balzer A, Jochim J, El CN, Fischer-Zirnsak B, Huber C, Kornak U, Elsea S, Cormier-Daire V, Ferreira C (2020). \u201cBiallelic variants in KYNU cause a multisystemic syndrome with hand hyperphalangism.\u201d Bone, 133, 115219. doi:10.1016/j.bone.2019.115219 \u261e Link
Seidel E, Walenda G, Messerschmidt C, Obermayer B, Peitzsch M, Wallace P, Bahethi R, Yoo T, Choi M, Schrade P, Bachmann S, Liebisch G, Eisenhofer G, Beule D, Scholl U (2020). \u201cGeneration and characterization of a mitotane-resistant adrenocortical cell line.\u201d Endocr Connect, 9, 122-134. doi:10.1530/EC-19-0510 \u261e Link
Melo U, Sch\u00f6pflin R, Acuna-Hidalgo R, Mensah M, Fischer-Zirnsak B, Holtgrewe M, Klever M, T\u00fcrkmen S, Heinrich V, Pluym I, Matoso E, Bernardo dSS, Louro P, H\u00fclsemann W, Cohen M, Dufke A, Latos-Biele\u0144ska A, Vingron M, Kalscheuer V, Quintero-Rivera F, Spielmann M, Mundlos S (2020). \u201cHi-C Identifies Complex Genomic Rearrangements and TAD-Shuffling in Developmental Diseases.\u201d Am J Hum Genet, 106, 872-884. doi:10.1016/j.ajhg.2020.04.016 \u261e Link
Robinson P, Ravanmehr V, Jacobsen J, Danis D, Zhang X, Carmody L, Gargano M, Thaxton C, UNC BC, Karlebach G, Reese J, Holtgrewe M, K\u00f6hler S, McMurry J, Haendel M, Smedley D (2020). \u201cInterpretable Clinical Genomics with a Likelihood Ratio Paradigm.\u201d Am J Hum Genet, 107, 403-417. doi:10.1016/j.ajhg.2020.06.021 \u261e Link
Blumenau S, Foddis M, M\u00fcller S, Holtgrewe M, Bentele K, Berchtold D, Beule D, Dirnagl U, Sassi C (2020). \u201cInvestigating APOE, APP-A\u03b2 metabolism genes and Alzheimer's disease GWAS hits in brain small vessel ischemic disease.\u201d Sci Rep, 10, 7103. doi:10.1038/s41598-020-63183-5 \u261e Link
Quade A, Thiel A, Kurth I, Holtgrewe M, Elbracht M, Beule D, Eggermann K, Scholl U, H\u00e4usler M (2020). \u201cParoxysmal tonic upgaze: A heterogeneous clinical condition responsive to carbonic anhydrase inhibition.\u201d Eur J Paediatr Neurol, 25, 181-186. doi:10.1016/j.ejpn.2019.11.002 \u261e Link
Thole T, Toedling J, Spr\u00fcssel A, Pfeil S, Savelyeva L, Capper D, Messerschmidt C, Beule D, Groeneveld-Krentz S, Eckert C, Gambara G, Henssen A, Finkler S, Schulte J, Sieber A, Bluethgen N, Regenbrecht C, K\u00fcnkele A, Lodrini M, Eggert A, Deubzer H (2020). \u201cReflection of neuroblastoma intratumor heterogeneity in the new OHC-NB1 disease model.\u201d Int J Cancer, 146, 1031-1041. doi:10.1002/ijc.32572 \u261e Link
Lamping M, Benary M, Leyvraz S, Messerschmidt C, Blanc E, Kessler T, Sch\u00fctte M, Lenze D, J\u00f6hrens K, Burock S, Klinghammer K, Ochsenreither S, Sers C, Sch\u00e4fer R, Tinhofer I, Beule D, Klauschen F, Yaspo M, Keilholz U, Rieke D (2020). \u201cSupport of a molecular tumour board by an evidence-based decision management system for precision oncology.\u201d Eur J Cancer, 127, 41-51. doi:10.1016/j.ejca.2019.12.017 \u261e Link
Praktiknjo S, Obermayer B, Zhu Q, Fang L, Liu H, Quinn H, Stoeckius M, Kocks C, Birchmeier W, Rajewsky N (2020). \u201cTracing tumorigenesis in a solid tumor model at single-cell resolution.\u201d Nat Commun, 11, 991. doi:10.1038/s41467-020-14777-0 \u261e Link
Suter A, Santos-Simarro F, Toerring P, Abad PA, Ramos-Mejia R, Heath K, Huckstadt V, Parr\u00f3n-Pajares M, Mensah M, H\u00fclsemann W, Holtgrewe M, Mundlos S, Kornak U, Bartsch O, Ehmke N (2020). \u201cVariable pulmonary manifestations in Chitayat syndrome: Six additional affected individuals.\u201d Am J Med Genet A, 182, 2068-2076. doi:10.1002/ajmg.a.61735 \u261e Link
Bockstahler M, Fischer A, Goetzke C, Neumaier H, Sauter M, Kespohl M, M\u00fcller A, Meckes C, Salbach C, Schenk M, Heuser A, Landmesser U, Weiner J, Meder B, Lehmann L, Kratzer A, Klingel K, Katus H, Kaya Z, Beling A (2020). \u201cHeart-Specific Immune Responses in an Animal Model of Autoimmune-Related Myocarditis Mitigated by an Immunoproteasome Inhibitor and Genetic Ablation.\u201d Circulation, 141, 1885-1902. doi:10.1161/CIRCULATIONAHA.119.043171 \u261e Link
Krishnamoorthy G, Kaiser P, Abu AU, Weiner J3, Moura-Alves P, Brinkmann V, Kaufmann S (2020). \u201cFX11 limits Mycobacterium tuberculosis growth and potentiates bactericidal activity of isoniazid through host-directed activity.\u201d Dis Model Mech, 13. doi:10.1242/dmm.041954 \u261e Link
Tregoning J, Weiner J, Cizmeci D, Hake D, Maertzdorf J, Kaufmann S, Leroux-Roels G, Maes C, Aerssens A, Calvert A, Jones C (2020). \u201cPregnancy has a minimal impact on the acute transcriptional signature to vaccination.\u201d NPJ Vaccines, 5, 29. doi:10.1038/s41541-020-0177-6 \u261e Link
Weiner J, Domaszewska T, Donkor S, Kaufmann S, Hill P, Sutherland J (2020). \u201cChanges in Transcript, Metabolite, and Antibody Reactivity During the Early Protective Immune Response in Humans to Mycobacterium tuberculosis Infection.\u201d Clin Infect Dis, 71, 30-40. doi:10.1093/cid/ciz785 \u261e Link
Win Z, Weiner J, Listanco A, Patel N, Sharma R, Greenwood A, Maertzdorf J, Mollenkopf H, Pizzoferro K, Cole T, Bodinham C, Kaufmann S, Denoel P, Del GG, Lewis D (2020). \u201cSystematic Evaluation of Kinetics and Distribution of Muscle and Lymph Node Activation Measured by (18)F-FDG- and (11)C-PBR28-PET/CT Imaging, and Whole Blood and Muscle Transcriptomics After Immunization of Healthy Humans With Adjuvanted and Unadjuvanted Vaccines.\u201d Front Immunol, 11, 613496. doi:10.3389/fimmu.2020.613496 \u261e Link
Mochel F, Rastetter A, Ceulemans B, Platzer K, Yang S, Shinde D, Helbig K, Lopergolo D, Mari F, Renieri A, Benetti E, Canitano R, Waisfisz Q, Plomp A, Huisman S, Wilson G, Cathey S, Louie R, Gaudio D, Waggoner D, Kacker S, Nugent K, Roeder E, Bruel A, Thevenon J, Ehmke N, Horn D, Holtgrewe M, Kaiser F, Kamphausen S, Abou JR, Weckhuysen S, Dalle C, Depienne C (2020). \u201cVariants in the SK2 channel gene (KCNN2) lead to dominant neurodevelopmental movement disorders.\u201d Brain, 143(12), 3564-3573. doi:10.1093/brain/awaa346 \u261e Link
Foddis M, Winek K, Bentele K, Mueller S, Blumenau S, Reichhart NN, Crespo-Garcia S, Harnett D, Ivanov A, Meisel A, Joussen A, Strauss O, Beule D, Dirnagl U, Sassi C (2020). \u201cAn exploratory investigation of brain collateral circulation plasticity after cerebral ischemia in two experimental C57BL/6 mouse models.\u201d J Cereb Blood Flow Metab, 40, 276-287. doi:10.1177/0271678X19827251 \u261e Link
Quade A, Thiel A, Kurth I, Holtgrewe M, Elbracht M, Beule D, Eggermann K, Scholl U, H\u00e4usler M (2020). \u201cParoxysmal tonic upgaze: A heterogeneous clinical condition responsive to carbonic anhydrase inhibition.\u201d Eur J Paediatr Neurol, 25, 181-186. doi:10.1016/j.ejpn.2019.11.002 \u261e Link
"},{"location":"misc/publication-list/#2019","title":"2019","text":"Kircher, M., Xiong, C., Martin, B. et al. \"Saturation mutagenesis of twenty disease-associated regulatory elements at single base-pair resolution.\" Nat Commun 10, 3583 (2019). doi: 10.1038/s41467-019-11526-w
Stefanovski L, Triebkorn P, Spiegler A, Diaz-Cortes M-A, Solodkin A, Jirsa V, McIntosh RA and Ritter P (2019). \"Linking Molecular Pathways and Large-Scale Computational Modeling to Assess Candidate Disease Mechanisms and Pharmacodynamics in Alzheimer's Disease.\" Front. Comput. Neurosci.. 13:54. doi: 10.3389/fncom.2019.00054
Boeddrich A., Babila J.T., Wiglenda T., Diez L., Jacob M., Nietfeld W., Huska M.R., Haenig C., Groenke N., Buntru A., Blanc E., Meier J.C., Vannoni E., Erck C., Friedrich B., Martens H., Neuendorf N., Schnoegl S., Wolfer DP., Loos M., Beule D., Andrade-Navarro M.A., Wanker E.E. (2019). \"The Anti-amyloid Compound DO1 Decreases Plaque Pathology and Neuroinflammation-Related Expression Changes in 5xFAD Transgenic Mice.\" Cell Chem Biol. 2019 Jan 17;26(1):109-120.e7. doi: 10.1016/j.chembiol.2018.10.013.
Fountain M.D., Oleson, D.S., Rech. M.E., Segebrecht, L., Hunter, J.V., McCarthy, J.M., Lupo, P.J., Holtgrewe, M., Mora, R., Rosenfeld, J.A., Isidor, B., Le Caignec, C., Saenz, M.S., Pedersen, R.C., Morgen, T.M., Pfotenhauer, J.P., Xia, F., Bi, W., Kang, S.-H.L., Patel, A., Krantz, I.D., Raible, S.E., Smith, W.E., Cristian, I., Tori, E., Juusola, J., Millan, F., Wentzensen, I.M., Person, R.E., K\u00fcry, S., B\u00e9zieau, S., Uguen, K., F\u00e9rec, C., Munnich, A., van Haelst, M., Lichtenbelt, K.D., van Gassen, K., Hagelstrom, T., Chawla, A., Perry, D.L., Taft, R.J., Jones, M., Masser-Frye, D., Dyment, D., Venkateswaran, S., Li, C., Escobar, L,.F., Horn, D., Spillmann, R.C., Pe\u00f1a, L., Wierzba, J., Strom, T.M. Parent, I. Kaiser, F.J., Ehmke, N., Schaaf, C.P. (2019). \"Pathogenic variants in USP7 cause a neurodevelopmental disorder with speech delays, altered behavior, and neurologic anomalies.\" Genet. Med. 2019 Jan 25. doi: 10.1038/s41436-019-0433-1
Holtgrewe,M., Messerschmidt,C., Nieminen,M. and Beule,D. (2019) DigestiFlow: from BCL to FASTQ with ease. Bioinformatics, 10.1093/bioinformatics/btz850.
K\u00e4fer S., Paraskevopoulou S., Zirkel F., Wieseke N., Donath A., Petersen M., Jones T.C., Liu S., Zhou X., Middendorf M., Junglen S., Misof B., Drosten C. (2019). \"Re-assessing the diversity of negative strand RNA viruses in insects.\" PLOS Pathogens 2019 Dec 12. doi: 10.1371/journal.ppat.1008224
K\u00fchnisch,J., Herbst,C., Al\u2010Wakeel\u2010Marquard,N., Dartsch,J., Holtgrewe,M., Baban,A., Mearini,G., Hardt,J., Kolokotronis,K., Gerull,B., et al. (2019) Targeted panel sequencing in pediatric primary cardiomyopathy supports a critical role of TNNI3. Clin Genet, 96, 549\u2013559. https://doi.org/10.1111/cge.13645
Marklewitz M., Dutari L.C., Paraskevopoulou S., Page R.A., Loaiza J.R., Junglen S. (2019). \"Diverse novel phleboviruses in sandflies from the Panama Canal area, Central Panama.\" Journal of General Virology 2019 May 3. doi: 10.1099/jgv.0.001260
Quade,A., Thiel,A., Kurth,I., Holtgrewe,M., Elbracht,M., Beule,D., Eggermann,K., Scholl,U.I. and H\u00e4usler,M. (2019) Paroxysmal tonic upgaze: A heterogeneous clinical condition responsive to carbonic anhydrase inhibition. European Journal of Paediatric Neurology, 10.1016/j.ejpn.2019.11.002.
Holtgrewe M, Messerschmidt C, Nieminen M, Beule D (2019). \u201cDigestiflow: From BCL to FASTQ with Ease.\u201d Bioinformatics.
Schumann F, Blanc E, Messerschmidt C, Blankenstein T, Busse A, Beule D (2019). \u201cSigsPack, a package for cancer mutational signatures.\u201d BMC Bioinformatics, 20, 450. doi:10.1186/s12859-019-3043-7 \u261e Link
Blanc E, Holtgrewe M, Dhamodaran A, Messerschmidt C, Willimsky G, Blankenstein T, Beule D (2019). \u201cIdentification and ranking of recurrent neo-epitopes in cancer.\u201d BMC Med Genomics, 12, 171. doi:10.1186/s12920-019-0611-7 \u261e Link
Boeddrich A, Babila J, Wiglenda T, Diez L, Jacob M, Nietfeld W, Huska M, Haenig C, Groenke N, Buntru A, Blanc E, Meier J, Vannoni E, Erck C, Friedrich B, Martens H, Neuendorf N, Schnoegl S, Wolfer D, Loos M, Beule D, Andrade-Navarro M, Wanker E (2019). \u201cThe Anti-amyloid Compound DO1 Decreases Plaque Pathology and Neuroinflammation-Related Expression Changes in 5xFAD Transgenic Mice.\u201d Cell Chem Biol, 26, 109-120.e7. doi:10.1016/j.chembiol.2018.10.013 \u261e Link
Borgsm\u00fcller N, Gloaguen Y, Opialla T, Blanc E, Sicard E, Royer A, Le BB, Durand S, Mign\u00e9 C, P\u00e9t\u00e9ra M, Pujos-Guillot E, Giacomoni F, Guitton Y, Beule D, Kirwan J (2019). \u201cWiPP: Workflow for Improved Peak Picking for Gas Chromatography-Mass Spectrometry (GC-MS) Data.\u201d Metabolites, 9. doi:10.3390/metabo9090171 \u261e Link
Buonfiglioli A, Efe I, Guneykaya D, Ivanov A, Huang Y, Orlowski E, Kr\u00fcger C, Deisz R, Markovic D, Fl\u00fch C, Newman A, Schneider U, Beule D, Wolf S, Dzaye O, Gutmann D, Semtner M, Kettenmann H, Lehnardt S (2019). \u201clet-7 MicroRNAs Regulate Microglial Function and Suppress Glioma Growth through Toll-Like Receptor 7.\u201d Cell Rep, 29, 3460-3471.e7. doi:10.1016/j.celrep.2019.11.029 \u261e Link
K\u00fchnisch J, Herbst C, Al-Wakeel-Marquard N, Dartsch J, Holtgrewe M, Baban A, Mearini G, Hardt J, Kolokotronis K, Gerull B, Carrier L, Beule D, Schubert S, Messroghli D, Degener F, Berger F, Klaassen S (2019). \u201cTargeted panel sequencing in pediatric primary cardiomyopathy supports a critical role of TNNI3.\u201d Clin Genet, 96, 549-559. doi:10.1111/cge.13645 \u261e Link
Kuhring M, Nieminen M, Kirwan J, Beule D, Holtgrewe M (2019). \u201cAltamISA: A Python API for ISA-Tab Files.\u201d JOSS, 4, 1610. doi:10.21105/joss.01610
Lamping M, Rieke D, Klauschen F, J\u00f6hrens K, Anagnostopoulos I, Lenze D, Tinhofer I, Benary M, Ochsenreither S, Klinghammer K, Burock S, Jann H, St\u00fcven A, Ditzen D, Beule D, Messerschmidt C, Blanc E, Sch\u00e4fer R, Keilholz U (2019). \u201cClinical Impact of Comprehensive versus Targeted Genomic Analysis for Precision Oncology.\u201d J Clin Oncol, 37(15_suppl), e13033-e13033. doi:10.1200/JCO.2019.37.15_suppl.e13033 \u261e Link
Herda S, Heimann A, Althoff S, Ru\u00df J, Bullinger L, Beule D, Obermayer B, Na I (2019). \u201cLong-Term T Cell Expansion Results in Increased Numbers of Central Memory T Cells with Sustained Functional Properties for Adoptive T Cell Therapy.\u201d Blood. doi:10.1182/blood-2019-124736 \u261e Link
van Heesch S, Witte F, Schneider-Lunitz V, Schulz J, Adami E, Faber A, Kirchner M, Maatz H, Blachut S, Sandmann C, Kanda M, Worth C, Schafer S, Calviello L, Merriott R, Patone G, Hummel O, Wyler E, Obermayer B, M\u00fccke M, Lindberg E, Trnka F, Memczak S, Schilling M, Felkin L, Barton P, Quaife N, Vanezis K, Diecke S, Mukai M, Mah N, Oh S, Kurtz A, Schramm C, Schwinge D, Sebode M, Harakalova M, Asselbergs F, Vink A, de WR, Viswanathan S, Widjaja A, G\u00e4rtner-Rommel A, Milting H, Dos RC, Knosalla C, Mertins P, Landthaler M, Vingron M, Linke W, Seidman J, Seidman C, Rajewsky N, Ohler U, Cook S, Hubner N (2019). \u201cThe Translational Landscape of the Human Heart.\u201d Cell, 178, 242-260.e29. doi:10.1016/j.cell.2019.05.010 \u261e Link
"},{"location":"misc/publication-list/#2018","title":"2018","text":"Blanc, E., Holtgrewe, M., Dhamodaran, A., Messerschmidt, C., Willimsky, G., Blankenstein, T., Beule, D. (2018). \"Identification and Ranking of Recurrent Neo-Epitopes in Cancer\". bioRxiv. 2018/389437, 2018. doi: 10.1101/389437
Brandt, R., Uhlitz, F., Riemer, P., Giesecke, C., Schulze, S., El-Shimy, I.A., Fauler, B., Mielke, T., Mages, N., Herrmann, B.G., Sers, C., Bl\u00fcthgen, N., Morkel, M. (2018). \"Cell type-dependent differential activation of ERK by oncogenic KRAS or BRAF in the mouse intestinal epithelium\". bioRxiv. 2018/340844. doi: 10.1101/340844.
Holtgrewe, M., Knaus, A., Hildebrand, G., Pantel, J.-T., Rodriguesz de los Santos, M., Neveling, K., Goldmann, J., Schubach, M., J\u00e4ger, M., Couterier, M., Mundlos, S., Beule, D., Sperling, K., Krawitz, P. (2018). \"Multisite de novo mutations in human offspring after paternal exposure to ionizing radiation\", Nature Scientific Reports. 2018 Oct 2;8(1):14611. doi: 10.1038/s41598-018-33066-x.
Kircher M., Xiong C., Martin B, Schubach M, Inoue F, Bell R.JA., Costello J.F., Shendure J., Ahituv N. (2018). \"Saturation mutagenesis of disease-associated regulatory elements.\" bioRxiv (2018): 505362. doi: 10.1101/505362
PCAWG Transcriptome Core Group, Calabrese, C., Davidson, N.R., Fonseca1, N.A., He, Y., Kahles, A., Lehmann, K.-V., Liu, F., Shiraishi, Y., Soulette, C.M., Urban, L., Demircio\u011flu, D., Greger, L., Li, S., Liu, D., Perry, M.D., Xiang, L., Zhang, F., Zhang, J., Bailey, P., Erkek, S., Hoadley, K.A., Hou, Y., Kilpinen, H., Korbel, J.O., Marin, M.G., Markowski, J., Nandi11, T., Pan-Hammarstr\u00f6m, Q., Pedamallu, C.S., Siebert, R., Stark, S.G., Su, H., Tan, P., Waszak, S.M., Yung, C., Zhu, S., PCAWG Transcriptome Working Group, Awadalla, P., Creighton, C.J., Meyerson, M., Ouellette, B.F.F., Wu, K., Yang, H., ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Network, Brazma1, A., Brooks, A.N., G\u00f6ke, J., R\u00e4tsch, G., Schwarz, R.F., Stegle, O., Zhang, Z. (2018). \"Genomic basis for RNA alterations revealed by whole-genome analyses of 27 cancer types\". bioRxiv. 2018/183889. doi: 10.1101/183889
Guneykaya D., Ivanov A., Hernandez D.P., Haage V., Wojtas B., Meyer N., Maricos M., Jordan P., Buonfiglioli A., Gielniewski B., Ochocka N., C\u00f6mert, C., Friedrich, C., Artiles, L. S., Kaminska, B., Mertins, P., Beule, D., Kettenmann, H. (2018). \"Transcriptional and translational differences of microglia from male and female brains\", Cell reports. 2018 Sep 4;24(10):2773-83. doi: 10.1016/j.celrep.2018.08.001.
Rentzsch P, Witten D, Cooper GM, Shendure J, Kircher M. (2018). \"CADD: predicting the deleteriousness of variants throughout the human genome\", Nucleic Acids Res. 2018 Oct 29. doi: 10.1093/nar/gky1016.
Salatzki J., Foryst-Ludwig A., Bentele K., Blumrich A., Smeir E., Ban Z., Brix S., Grune J., Beyhoff N., Klopfleisch R., Dunst S., Surma, M.A., Klose, C., Rothe, M., Heinzel, F.R., Krannich, A., Kershaw, E.E., Beule, D., Schulze, P.C., Marx, N., Kintscher, U. (2018). \"Adipose tissue ATGL modifies the cardiac lipidome in pressure-overload-induced left ventricular failure\", PLoS genetics. 2018 Jan 10;14(1):e1007171. doi: 10.1371/journal.pgen.100717.
Schubach M., Re M., Robinson P.N., Valentini G. (2017) \"Imbalance-aware machine learning for predicting rare and common disease-associated non-coding variants\", Scientific reports 7:1, 2959. doi: 10.1038/s41598-017-03011-5.
Schubert M., Klinge, B., Kl\u00fcnemann M., Sieber A., Uhlitz F., Sauer S., Garnett M., Bl\u00fcthgen N., Saez-Rodriguez J. (2018). \"Perturbation-response genes reveal signaling footprints in cancer gene expression\". Nature Communications. 9: 20, 2018. doi: 10.1038/s41467-017-02391-6
"},{"location":"misc/publication-list/#2017","title":"2017","text":"Euskirchen, P., Bielle, F., Labreche, K., Kloosterman, W.P., Rosenberg, S., Daniau, M., Schmitt, C., Masliah-Planchon, J., Bourdeaut, F., Dehais, C., et al. (2017). Same-day genomic and epigenomic diagnosis of brain tumors using real-time nanopore sequencing. Acta Neuropathol 1\u201313. doi: 10.1007/s00401-017-1743-5
Euskirchen, P., Radke, J., Schmidt, M.S., Heuling, E.S., Kadikowski, E., Maricos, M., Knab, F., Grittner, U., Zerbe, N., Czabanka, M., et al. (2017). Cellular heterogeneity contributes to subtype-specific expression of ZEB1 in human glioblastoma. PLOS ONE 12, e0185376. doi: 10.1371/journal.pone.0185376
Mattei D., Ivanov A., Ferrai C., Jordan P., Guneykaya D., Buonfiglioli A., Schaafsma W., Przanowski P., Deuther-Conrad W., Brust P., Hesse S., Patt, M., Sabri, O., Ross, T.L., Eggen, B.J.L., Boddeke E.W.G.M., Kaminska, B., Beule, D., Pombo, A., Kettenmann, H., Wolf, S.A. (2017). \"Maternal immune activation results in complex microglial transcriptome signature in the adult offspring that is reversed by minocycline treatment.\" Translational psychiatry. 2017 May;7(5):e1120. doi: 10.1038/tp.2017.80.
Mamlouk, S., Childs, L. H., Aust, D., Heim, D., Melching, F., Oliveira, C., Wolf, T., Durek, P., Schumacher, D., Bl\u00e4ker, H., von Winterfeld, M., Gastl, B., M\u00f6hr, K., Menne, A., Zeugner, S., Redmer, T., Lenze, D., Tierling, S., M\u00f6bs, M., Weichert, W., Folprecht, G., Blanc, E., Beule, D., Sch\u00e4fer, R., Morkel, M., Klauschen, F., Leser, U. and Sers, C. (2017). \"DNA copy number changes define spatial patterns of heterogeneity in colorectal cancer\", Nature Communications. 2017; 8, p. 14093. doi: 10.1038/ncomms14093.
Messerschmidt, C., Holtgrewe, M. and Beule, D. (2017). \"HLA-MA: simple yet powerful matching of samples using HLA typing results\". Bioinformatics. 28, pp. 2592\u20132599. doi: 10.1093/bioinformatics/btx132.
Kammertoens, T., Friese, C., Arina, A., Idel, C., Briesemeister, D., Rothe, M., Ivanov, A., Szymborska, A., Patone, G., Kunz, S., Sommermeyer, D., Engels, B., Leisegang, M., Textor, A., Fehling, H. J., Fruttiger, M., Lohoff, M., Herrmann, A., Yu, H., Weichselbaum, R., Uckert, W., H\u00fcbner, N., Gerhardt, H., Beule, D., Schreiber, H. and Blankenstein, T. (2017). \"Tumour ischaemia by interferon-\u03b3 resembles physiological blood vessel regression\". Nature. 545(7652), pp. 98\u2013102. doi: 10.1038/nature22311.
Schulze Heuling, E., Knab, F., Radke, J., Eskilsson, E., Martinez-Ledesma, E., Koch, A., Czabanka, M., Dieterich, C., Verhaak, R.G., Harms, C., et al. (2017). Prognostic Relevance of Tumor Purity and Interaction with MGMT Methylation in Glioblastoma. Mol. Cancer Res. 15, 532\u2013540. doi: 10.1158/1541-7786.MCR-16-0322
Yaakov, G., Lerner, D., Bentele, K., Steinberger, J., Barkai, N., Bigger, J., Maisonneuve, E., Gerdes, K., Lewis, K., Dhar, N., McKinney, J. D., Gefen, O., Balaban, N. Q., Jayaraman, R., Balaban, N. Q., Merrin, J., Chait, R., Kowalik, L., Leibler, S., Balaban, N. Q., Allison, K. R., Brynildsen, M. P., Collins, J. J., Nathan, C., Lewis, K., Glickman, M. S., Sawyers, Knoechel, B., Welch, A. Z., Gibney, P. A., Botstein, D., Koshland, D. E., Levy, S. F., Ziv, N., Siegal, M. L., Stewart-Ornstein, J., Weissman, J. S., El-Samad, H., Gasch, A. P., Weinert, T., Hartwell, L., Weinert, T. A., Hartwell, L. H., Lisby, M., Rothstein, R., Mortensen, U. H., Lisby, M., Mortensen, U. H., Rothstein, R., Domkin, V., Thelander, L., Chabes, A., Hendry, J. A., Tan, G., Ou, J., Boone, C., Brown, G. W., Berry, D. B., Gasch, A. P., Lynch, M., Nishant, K. T., Serero, A., Jubin, C., Loeillet, S., Legoix-Ne, P., Nicolas, A. G., Huh, W. K., Janke, C., Lee, S. E., Blecher-Gonen, R., Martin, M., Cherry, J. M., McKenna, A., DePristo, M. A., Lawrence, M., Obenchain, V., Ye, K., Schulz, M. H., Long, Q., Apweiler, R., Ning, Z., Layer, R. M., Chiang, C., Quinlan, A. R., Hall, I. M., Faust, G. G., Hall, I. M., Boeva, V., Boeva, V., Li, H., Koren, A., Soifer, I. and Barkai, N. (2017). \"Coupling phenotypic persistence to DNA damage increases genetic diversity in severe stress\". Nature Ecology & Evolution. 1(1), pp. 497\u2013500. doi: 10.1038/s41559-016-0016.
Uhlitz, F., Sieber, A., Wyler, E., Fritsche-Guenther, R., Meisig, J., Landthaler, M., Klinger, B., Bl\u00fcthgen, N. (2017). \"An immediate-late gene expression module decodes ERK signal duration\". Molecular Systems Biology. 13: 928, 2017. doi: 10.15252/msb.20177554.
"},{"location":"misc/publication-list/#theses","title":"Theses","text":""},{"location":"misc/publication-list/#2019_1","title":"2019","text":"Schumann F. (2019). \"Establishing a pipeline for stable mutational signature detection and evaluation of variant filter effects\". Freie Universit\u00e4t Berlin. Bachelor Thesis, Bioinformatics.
"},{"location":"misc/publication-list/#2018_1","title":"2018","text":"Borgsm\u00fcller N. (2018). \"Optimization of data processing in GC-MS metabolomics\", Technische Universit\u00e4t Berlin. Master Thesis, Biotechnology.
Kuchenbecker, S.-L. (2018). \"Analysis of Antigen Receptor Repertoires Captured by High Throughput Sequencing\". Freie Universit\u00e4t Universit\u00e4t Berlin. PhD Thesis, Dr. rer. nat. URN:NBN: urn:nbn:de:kobv:188-refubium-22171-8
Schubach M. (2018). \"Learning the Non-Coding Genome\", Freie Universit\u00e4t Universit\u00e4t Berlin. PhD Thesis, Dr. rer. nat. URN:NBN: urn:nbn:de:kobv:188-refubium-23332-7
"},{"location":"misc/publication-list/#posters","title":"Posters","text":""},{"location":"misc/publication-list/#2018_2","title":"2018","text":"Roskosch, S., Hald\u00f3rsson B., Kehr, B. (2018). \"PopDel: Population-Scale Detection of Genomic Deletions\" ECCB 2018. Poster.
White T., Kehr B. (2018). \"Comprehensive extraction of structural variations from long-read DNA sequences\" WABI 2018. Poster.
"},{"location":"misc/publication-list/#2017_1","title":"2017","text":"Schubach M., Re R., Robinson P.N., Valentini G. (2017). \"Variant relevance prediction in extremely imbalanced training sets\" ISMB/ECCB 2017. Poster.
White T., Kehr B. (2017). \"Improving long-read mapping with simple lossy sequence transforms\" ISMB/ECCB 2017. Poster.
"},{"location":"ondemand/interactive/","title":"OnDemand: Interactive Sessions","text":"Interactive sessions allow you to start and manage selected apps. Depending on the app they run as servers or GUIs. Selecting My Interactive Sessions in the top menu will direct you to the overview of currently running sessions. The left-hand panel provides a short cut to start a new session of one of the provided apps.
Each running interactive session is listed. Each card corresponds to one session. The title of each card provides the name, allocated resources and the current status. Furthermore, detailed information and links are available:
Don't hit reload in your apps
Please note that the portal will use the authentication mechanisms of the apps to ensure that nobody except for you can connect to the session. This means that hitting the browsers \"reload\" button in your app will most likely not work.
Just go back to the interactive session list and click on the \"connect\" button.
"},{"location":"ondemand/interactive/#session-directories","title":"Session Directories","text":"The portal software will create a folder ondemand in your home directory. Inside, it will create session directories for each started interactive job. For technical reasons, these folders have very long names, for example:
$HOME/ondemand/data/sys/dashboard/batch_connect/sys/ood-bih-rstudio-server/output/e40e03b3-11ca-458a-855b-98e6f148c99a/This follows the pattern:
$HOME/${application name}/output/${job UUID}The job identifier used is not the Slurm job ID but an identifier internal to OnDemand. Inside this directory you will find log files and a number of scripts that are used to start your job.
If you need to debug any interactive job, start here. Also, the helpdesk will need the path to this folder to help you with interactive jobs.
You can find the name of the latest output folder with the following command:
$ ls -lhtr $HOME/${application name}/output | tail -n 1\nFor example, for RStudio Server:
$ ls -lhtr $HOME/ondemand/data/sys/dashboard/batch_connect/sys/ood-bih-rstudio-server/output | tail -n 1\nPrevent Home From Filling Up
You should probably move ~/ondemand to your work volume with the following:
$ mv ~/ondemand ~/work/ondemand\n$ ln -sr ~/work/ondemand ~/ondemand\nMake sure to delete potential interactive sessions and to logout from the Ondemand Portal first. Otherwise, the ~/ondemand folder is constantly recreated and the symlink will be just created within this folder as ~/ondemand/ondemand and thus not be used as intended.
Also, clear out ~/work/ondemand/* from time to time but take care that you don't remove the directory of any running job.
This description of starting an RStudio session is a showcase for starting other interactive apps as well.
To start the session, please go to Interactive Apps in the top menu bar and select RStudio Server or click RStudio Server in the left-hand panel.
Allocate appropriate resources and click Launch.
An info card for the RStudio Server will be added to My Interactive Sessions, and during start, it will change its state from Queued to Starting to Running. Depending on the app, resources allocated and current cluster usage, this will take a couple of seconds.
When in the final state (Running), one can directly connect to the RStudio Server to get an interactive session by clicking Connect to RStudio Server:
To use the OnDemand portal with a specific R installation including a stable set of custom packages you can use a conda enviroment from the cluster as a R source.
For this you may first need to create this conda environment including your R version of choice and all necessary packages. Specific installations of i.e. python from conda can be used similarly in other interactive apps.
channels:\n - conda-forge\n - bioconda\n - defaults\ndependencies:\n - r-base\n - r-essentials\n - r-devtools\n - bioconductor-deseq2\n - r-tidyverse\n - r-rmarkdown\n - r-knitr\n - r-dt\nSome packages (i.e. several single-cell-RNAseq analysis tools) are only available from github and not on Cran/Bioconductor. There are two ways to install such packages into a conda enviroment.
Click to expand 1) Install from inside R \\[easier option, but not pure conda\\] * First setup the conda env, ideally including all dependencies for the desired package from github (and do include r-devtools) * Then within R run `devtools::install_github('owner/repo', dependencies=F, upgrade=F, lib='/path/to/conda/env-name/lib/R/library')` * if you don't have all dependencies already installed you will have to omit dependencies=F and risk a mix of conda & native R installed packages (or just have to redo the conda env). * github_install involves a build process and still needs a bit of memory, so this might crash on the default `srun --pty bash -i` shell 2) Build packages into a local conda channel \\[takes longer, but pure conda\\]\\ This approach is mostly taken from the answers given [here](https://stackoverflow.com/questions/52061664/install-r-package-from-github-using-conda). These steps must be taken _before_ building the final env used with Rstudio * use `conda skeleton cran https://github.com/owner/repo [--git-tag vX.Y]` to generate build files * conda skeleton only works for repositories with a release/version tag. If the package you want to install does not have that, you either need to create a fork and add a such a tag, or find a fork that already did that. Downloading the code directly from github and building the package from that is also possible, but you will the need to manually set up the `meta.yaml` and `build.sh` files that conda skeleton would create. * If there is more than one release tag, do specify which one you want, it may not automatically take the most recent one. * If any r-packages from bioconductor are dependencies, conda will not find them during the build process. You will need to change the respective entries in the `meta.yaml` file created by conda skeleton. I.e. change `r-deseq2` to `bioconductor-deseq2` * Build the package with `conda build --R= [--use-local] r-` * You need to specifying the same R-version used in the final conda env * If the github package has additional dependencies from github, build those first and then add `--use-local` so the build process can find them. * The build process definitely needs more memory than the default `srun --pty bash -i` shell. It also takes quite a bit of time (much longer than installing through devtools::install_github) * Finally add the packages (+versions) you built to the environment definition (i.e. yaml file) and create the (final) conda environment. Don't forget to tell conda to use locally build packages (either supply `--use-local` or add `- local` to the channel list in the yaml file)Starting the Rstudio session via the OnDemand portal works almost as described above (see Example 1). However, you do have to select `miniforge` as R source and provide the path to your miniforge installation and (separated by a colon) the name of the (newly created) conda enviroment you want to use.
Additional notes:
.libPaths() entries and therefore a link to your previous conda installation. Creating a new project cleans .libPaths() to only the env specified in setting up the Rstudio session.Status / Stability
OnDemand Support is currently in beta phase on the BIH HPC. In case of any issues, please send an email to hpc-helpdesk@bih-charite.de.
To allow for better interactive works, BIH HPC administration has setup an Open OnDemand (OOD) portal web server.
You can find the OnDemand Portal for HPC 4 Research at:
OOD allows you to access cluster resources using a web-based graphical interface in addition to traditional SSH connections. You can then connect to jobs running graphical applications either to virtual desktops (such as Matlab) or to web apps (such as Jupyter and RStudio Server).
The following figure illustrates this.
The primary way to the cluster continues to be SSH which has several advantages. By the nature of the cluster being based on Linux servers, it will offer more features through the \"native\" access and through its lower complexity, it will offer higher stability. However, we all like to have the option of a graphical interface, at least from time to time .
The main features are:
The first prerequisite is to have a cluster account already (see Getting Access). Once you have done your first SSH connection to the cluster successfully you can start using the portal. For this you perform the following steps:
_c) then please use the \"Charit\u00e9 - Universit\u00e4tmedizin Berlin\" button, for MDC Accounts please use the \"Max Delbr\u00fcck Center Berlin\" button. Clicked the Wrong Login Button?
If you clicked the wrong button then please clear your cookies to force a logout of the system.
"},{"location":"ondemand/overview/#prepare-ondemand-folder","title":"Prepare OnDemand Folder","text":"The ondemand folder is automatically created in your home directory, and the OnDemand service searches for this folder in your home directory, i.e. it has to stay there. But as the quota in the home directory is very limited, you can easily hit the hard quota which might prevent you from working on the cluster.
To prevent this, move the ~/ondemand folder to the ~/work folder and create a symlink for the now dislocated ~/ondemand folder:
hpc-login-1:~$ mv ~/ondemand ~/work/ondemand\nhpc-login-1:~$ ln -sr ~/work/ondemand ~/ondemand\nImportant
Make sure to delete potential interactive sessions and to logout from the Ondemand Portal first. Otherwise, the ~/ondemand folder is constantly recreated and the symlink will be just created within this folder as ~/ondemand/ondemand and thus not be used as intended.
Problems with Open OnDemand?
First try to log out and login again. Next, try to clear all cookies for the domain hpc-portal.cubi.bihealth.org. Finally, try the Help > Restart Web Server link to restart the per-user nginx (PUN) server.
You will then be redirected to the dashboard screen.
Here you have access to the following actions. We will not go into detail of all of them and expect them to be self-explanatory.
Important
Please note that when using the portal then you are acting as your HPC user. Use standard best practice. Consider carefully what you do as you would from the command line (e.g., don't use the portal to browse the web from the cluster).
Outdated
This document is only valid for the old, third-generation file system and will be removed soon. Quotas of our new CephFS storage are communicated via the HPC Access web portal.
Accessing the quota report by selecting Files and then Quotas in the top menu will provide you with a detailed list of all quotas for directories that you are assigned to.
There are two types of quotas: for (a) size of and (b) number of files in a directory.
Every row in the table corresponds to a directory that you have access to. This implies your home directory (fast/users) as well as the group directory of your lab (fast/groups) and possible projects (fast/projects) (if any). Quotas are not directly implied on these directories but on the home, scratch and work subdirectories that each of subdirectory of the beforementioned directories has (for a detailed explanation see Storage and Volumes).
The following list explains the columns of the table:
/) and substituting the underscores with a slash in the (users|groups|projects)_ and _(home|scratch|work) substring. The corresponding path for name fast/users_stolpeo_c_home would be /fast/users/stolpeo_c/home.BIH HPC IT provides acess to high-performance compute (HPC) cluster systems. A cluster system bundles a high number of nodes and in the case of HPC, the focus is on performance (with contrast to high availability clusters).
"},{"location":"overview/architecture/#hpc-4-research","title":"HPC 4 Research","text":""},{"location":"overview/architecture/#cluster-hardware","title":"Cluster Hardware","text":"Users don't connect to nodes directly but rather create interactive or batch jobs to be executed by the cluster job scheduler Slurm.
As common with HPC systems, users cannot directly access the compute nodes but rather connect to so-called head nodes. The BIH HPC system provides the following head nodes:
login-1 and login-2 that accept SSH connections and are meant for low intensity, interactive work such as editing files, running screen/tmux sessions, and logging into the compute nodes. Users should run no computational tasks and no large-scale data transfer on these nodes.transfer-1 and transfer-2 also accept SSH connections. Users should run all large-scale data transfer through these nodes.After registration and client configurations, users with typically connect to the HPC system through the login nodes:
local:~$ ssh -l jdoe_c hpc-login-1.cubi.bihealth.org\nhpc-login-1:~$\nSubsequently, they might submit batch jobs to the cluster for execution through the Slurm scheduling system or open interactive sessions:
hpc-login-1:~$ sbatch job_script.sh\nhpc-login-1:~$ srun --pty bash -i\nmed0104:~$\nBIH HPC 4 Research is located in the BIH data center in Buch and connected via the BIH research network. Connections can be made from Charite, MDC, and BIH networks. The cluster is open for users with either Charite or MDC accounts after getting access through the gatekeeper proces. The system has been designed to be suitable for the processing of human genetics data from research contexts (and of course data without data privacy concerns such as public and mouse data).
"},{"location":"overview/for-the-impatient/#cluster-hardware-and-scheduling","title":"Cluster Hardware and Scheduling","text":"The cluster consists of the following major components:
hpc-login-1 and hpc-login-2 (for interactive sessions only),hpc-transfer-1 and hpc-transfer-2,hpc-cpu-{1..228}hpc-mem-{1..5},hpc-gpu-{1..7} and 1 node with 10x A40 GPUs (!) hpc-gpu-8,/fast,This is shown by the following picture:
"},{"location":"overview/for-the-impatient/#differences-between-workstations-and-clusters","title":"Differences Between Workstations and Clusters","text":"The differences include:
srun to go to a compute node.srun to go to a compute node you might end up on a different host./tmp./fast directory is shared throughout the cluster which contains your home, group home, and project directories.root or sudo permissions on the cluster.sbatch) over calling programs interactively.NB: the following might sound a bit harsh but is written with everyone's best intentions in mind (we actually like you, our user!) This addresses a lot of suboptimal (yet not dangerous, of course) points we observed in our users.
IT IS
IT IS NOT
sudo.Once logged into the cluster through the login nodes, the Slurm scheduler needs to be used to submit computing jobs. In Slurm nomenclature, cluster compute nodes are assigned to one or more partitions. Submitted jobs are assigned to nodes according to the partition's configuration.
"},{"location":"overview/job-scheduler/#partitions","title":"Partitions","text":"The BIH HPC has the partitions described below. The cluster focuses on life science applications and not \"classic HPC\" with numerical computations using MPI. Thus, all partitions except for mpi only allow to reserve resources on one node. This makes the cluster easier to use as users don't have to explicitely specify this limit when submitting their jobs.
standard","text":"Jobs are submitted to the standard partition by default. From the, the scheduler will route the jobs to their actual partition using the routing rule set described below. You can override this routing by explicitely assigning a partition (but this is discouraged).
gpu queue.highmem queue.debug, short, medium, and long long depending on their configured maximal running time. The partitions are evaluated in the order given above and the first fitting partition will be used.debug","text":"This partition is for very short jobs that should be executed quickly, e.g., for tests. The job running time is limited to one hour and at most 128 cores can be used per user but the jobs are submitted with highest priority.
debug--time 01:00:00short","text":"This partition is for jobs running only few hours. The priority of short jobs is high and many cores can be used at once to reward users for splitting their jobs into smaller parts.
short--time 04:00:00medium","text":"This partition is for jobs running for multiple days. Users can only allocate the equivalent of 4 nodes.
medium--time 7-00:00:00long","text":"This partition is for long-running tasks. Only one node can be reserved for so long to discourage really long-running jobs and encourage users for splitting their jobs into smaller parts.
long--time 14-00:00:00gpu","text":"Jobs requesting GPU resources are automatically assigned to the gpu partition.
The GPU nodes are only part of the gpu partition so they are not blocked by normal compute jobs. Maximum run time is relatively high (14 days) to allow for longer training jobs. Contact hpc-helpdesk@bih-charite.de if you have longer running jobs that you really cannot make run any shorter for assistance.
Info
Fair use rules apply. As GPU nodes are a limited resource, excessive use by single users is prohibited and can lead to mitigating actions. Be nice and cooperative with other users. Tip: getent passwd USER_NAME will give you a user's contact details.
gpu$count GPUs: -p gpu --gres=gpu:$card:$count (card=tesla or card=a40), maximum run time: --time 14-00:00:00highmem","text":"Jobs requesting more than 200 GB of RAM are automatically routed to the highmem partition.
The high memory nodes are only part of the highmem partition so they are not blocked by normal compute jobs. Maximum run time is relatively high (14 days) to allow for longer jobs. Contact hpc-helpdesk@bih-charite.de for assistance if you have longer running jobs that you really cannot make run any shorter.
Info
Fair use rules apply. As high-memory nodes are a limited resource, excessive use by single users is prohibited and can lead to mitigating actions. Be nice and cooperative with other users. Tip: getent passwd USER_NAME will give you a user's contact details.
highmem-p highmem, maximum run time: --time 14-00:00:00mpi","text":"Jobs are not routed automatically to the mpi partition but you have to explitely request the partition. This is the only partition in which more than one node can be allocated to a job.
You can submit multi-node jobs into the mpi partition. Maximum run time is relatively high (14 days) to allow for longer jobs. Don't abuse this. Contact hpc-helpdesk@bih-charite.de for assistance if you have longer running jobs that you really cannot make run any shorter.
highmem-p mpi, maximum run time: --time 14-00:00:00critical","text":"Jobs are not routed into critial automatically and the partition has to be selected manually.
This partition is for time-critical jobs with deadlines. As long as the cluster is not very busy, requests for critical jobs will be granted most of the time. However, do not use this partition without arranging with hpc-helpdesk as killing jobs will be used as the ultima ratio in case of such policy violations.
critical--time 7-00:00:00We currently provide a Grafana dashboard for monitoring various aspects of the cluster's status:
https://metrics.cubi.bihealth.org/public-dashboards/dc3e4d5b1ea049429abf39e412c47302
"},{"location":"overview/storage/","title":"Nodes and Storage Volumes","text":"No mounting on the cluster itself.
For various technical and security-related reasons it is not possible to mount anything on the cluster nodes by users. For mounting the cluster storage on your computer, please read Connecting: SSHFS Mounts.
This document gives an overview of the nodes and volumes on the cluster.
"},{"location":"overview/storage/#cluster-layout","title":"Cluster Layout","text":""},{"location":"overview/storage/#cluster-nodes","title":"Cluster Nodes","text":"The following groups of nodes are available to cluster users. There are a number of nodes that are invisible to non-admin staff, hosting the queue master and monitoring tools and providing backup storage for key critical data, but these are not shown here.
hpc-login-{1,2}hpc-login-{1,2}.cubi.bihealth.orgmed0101..0124,0127 med0133..0164med0201..0264med0301..0304med0401..0405 special purpose/high-memory machinesmed0401 and med0402med0403 and med0404med0405gpu)med0601..0616med0618..0633med0701..0764The cluster has 2.1 PB of legacy fast storage, currently available at /fast, as well as 1.6 PB of next-generation fast storage, available at /data/cephfs-1. Additionally 7.4 PB of slower \"Tier 2\" storage is available at /data/cephfs-2. Storage is provided by a Ceph storage cluster and designed for massively parallel access from an HPC system. In contrast to \"single server\" NFS systems, the system can provide large bandwidth to all cluster nodes in parallel as long as large data means relatively \"few\" files are read and written.
Storage is split into three sections:
home -- small, persistent, and safe storage, e.g., for documents and configuration files (default quota of 1 GB).work -- larger and persistent storage, e.g., for your large data files (default quota of 1 TB).scratch -- large and non-persistent storage, e.g., for temporary files, files are automatically deleted after 2 weeks (default quota of 10 TB; deletion not implemented yet).)Each user, group, and project has one or more of these sections each, e. g. for users:
/data/cephfs-1/home/users/$NAME/data/cephfs-1/home/users/$NAME/work/data/cephfs-1/home/users/$USER/scratchSee Storage and Volumes: Locations for more informatin.
"},{"location":"slurm/background/","title":"Introduction to Scheduling","text":"As explained elsewhere in more detail, an HPC cluster consists of multiple computers connected via a network and working together. Multiple users can use the system simultaneously to do their work. This means that the system needs to join multiple computers (nodes) to provide a coherent view of them and the same time partition the system to allow multiple users to work concurrently.
user 1 user 2 ...\n\n .---. .---. .---. .---.\n | J | | J | | J | | J |\n | o | | o | | o | | o | ...\n | b | | b | | b | | b |\n | 1 | | 2 | | 3 | | 4 |\n '---' '---' '---' '---'\n\n.------------------------------------------.\n| Cluster Scheduler |\n'------------------------------------------'\n\n.----------. .------------. .------------.\n| multiple | | separate | | computers |\n'----------' '------------' '------------'\nOverall, this partitioning is not so different from how your workstation or laptop works. Most likely, your computer (or even your smartphone) has multiple processors (or cores). You can run multiple programs on the same computer and the fact that (a) there is more than one core and (b) there is more than one program running is not known to the running programs (unless they explicitly communicate with each other). Different programs can explicitly take advantage of the multiple processor cores. The main difference is that you normally use your computer in an interactive fashion (you perform an action and expect an immediate reaction).
Even with a single processor (and core), your computer manages to run more than one program at the same time. This is done with the so-called time-slicing approach where the operating system lets each programs run in turn for a short time (a few milliseconds). A program with a higher priority will get more time slices than one with a lower (e.g., your audio player has real-time requirements and you will hear artifacts if it is starved for compute resources). Your operating system protects programs from each other by creating an address space for each. When two programs are running, the value of the memory at any given position in one program is independent from the value in the other program. Your operating system offers explicit functionality for sharing certain memory areas that two programs can use to exchange data efficiently.
Similarly, file permissions with Unix users/groups or Unix/Windows ACLs (access control lists) are used to isolate users from each other. Programs can share data by accessing the same file if they can both access it. There are special files called sockets that allow for network-like inter-process communication but of course two programs on the same computer can also connect (virtually) via the computer network (no data will actually go through a cable).
"},{"location":"slurm/background/#interlude-resource-types","title":"Interlude: Resource Types","text":"As another diversion, let us consider how Unix manages its resources. This is important to understand when requesting resources from the scheduler later on.
First of all, a computer might offer a certain feature such as a specific hardware platform or special network connection. Examples for this on the BIH HPC are specific Intel processor generations such as haswell or the availability of Infiniband networking. You can request these with so-called constraints; they are not allocated to specific jobs.
Second, there are resources that are allocated to specific jobs. The most important resources here are:
Generally, once a resource has been allocated to one job, it is not available to another. This means if you allocating more resources to your job that you actually need (overallocation) then those resources are not available to other jobs (whether they are your jobs or those of other users). This will be explained further below.
Another example of resource allocation are licenses. The BIH HPC has a few Matlab 2016b licenses that users can request. As long as a license is allocated to one job, it is unavailable to another.
"},{"location":"slurm/background/#nodes-sockets-processors-cores-threads","title":"Nodes, Sockets, Processors, Cores, Threads","text":"Regarding compute resources, Slurm differentiates between:
In most cases, you will use one compute node only. When using more than one node, you will need to use some form of message passing, e.g., MPI, so processes on different nodes can communicate. On a single node you would mostly use single- or multi-threaded processes, or multiple processes.
Above: Slurm's nomenclature for sockets, processors, cores, and threads (from Slurm Documentation).
Co-locating processes/threads on the same socket has certain implications that are mostly useful for numerical applications. We will not further go into detail here. Slurm provides many different features of ways to specify allocation of \"pinning\" to specific process locations. If you need this feature, we trust that you find sufficient explanation in the Slurm documentation.
Usually, you would allocate multiple cores (a term Slurm uses synonymously with processors) on a single node (allocation on a single node is the default).
"},{"location":"slurm/background/#how-scheduling-works","title":"How Scheduling Works","text":"Slurm is an acronym for \"Simple Linux Unix Resource Manager\" (note that the word \"scheduler\" does not occur here). Actually, one classically differentiates between the managing of resources and the scheduling of jobs that use them. The resource manager allocates resources according to a user's request for a job and ensures that there are no conflicts. If the required resources are not available, the scheduler puts the user's job into a queue. Later, when then requested resources become available the scheduler assigns them to the job and runs it. In the following, both resource allocation and the running of the job are described as being done by the scheduler.
The interesting case occurs when there are not enough resources available for at least two jobs submitted to the scheduler. The scheduler has to decide how to proceed. Consider the simplified case of only scheduling cores. Each job will request a number of cores. The scheduler will then generate a scheduling plan that might look as follows.
core\n ^\n4 | |---job2---|\n3 | |---job2---|\n2 | |---job2---|\n1 | |--job1--|\n +--------------------------> t time\n 5 1 1 2\n 0 5 0\njob1 has been allocated one core and job2 has been allocated two cores. When job3, requesting one core is submitted at t = 5, it has to wait at least as long until job1 is finished. If job3 requested two or more cores, it would have to wait at least until job2 also finished.
We can now ask several questions, including the following:
Also see the Slurm Frequently Asked Questions.
Please note that even if all jobs were known at the start of time, scheduling is still a so-called NP-complete problem. Entire computer science journals and books are dedicated only to scheduling. Things get more complex in the case of online scheduling, in which new jobs can appear at any time. In practice, Slurm does a fantastic job with its heuristics but it heavily relies on parameter tuning. HPC administration is constantly working on optimizing the scheduler settings. Note that you can use the --format option to the squeue command to request that it shows you information about job scheduling (in particular, see the %S field, which will show you the expected start time for a job, assuming Slurm has calculated it). See man squeue for details. If you observe inexplicable behavior, please notify us at hpc-helpdesk@bih-charite.de.
In Slurm, the nodes of a cluster are split into partitions. Nodes are assigned to one or more partition (see the Job Scheduler section for details). Jobs can also be assigned to one or more partitions and are executed on nodes of the given partition.
In the BIH HPC, partitions are used to stratify jobs of certain running times and to provide different quality of service (e.g., maximal number of CPU cores available to a user for jobs of a certain running time and size). The partitions gpu and highmem provide special hardware (the nodes are not assigned to other partitions) and the mpi partition allows MPI-parallelism and the allocation of jobs to more than one node. The Job Scheduler provides further details.
This page contains assorted Slurm commands and Bash snippets that should be helpful.
man pages!
$ man sinfo\n$ man scontrol\n$ man squeue\n# etc...\ninteractive sessions
hpc-login-1:~$ srun --pty bash\nmed0740:~$ echo \"Hello World\"\nmed0740:~$ exit\nbatch submission
hpc-login-1:~$ sbatch script.sh\nSubmitted batch job 2\nhpc-login-1:~$ squeue\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 27 debug script.s holtgrem R 0:06 1 med0703\nlisting nodes
$ sinfo -N\nNODELIST NODES PARTITION STATE\nmed0740 1 debug* idle\nmed0741 1 debug* down*\nmed0742 1 debug* down*\n\n$ scontrol show nodes\nNodeName=med0740 Arch=x86_64 CoresPerSocket=8\n CPUAlloc=0 CPUTot=32 CPULoad=0.06\n AvailableFeatures=(null)\n[...]\n\n$ scontrol show nodes med0740\nNodeName=med0740 Arch=x86_64 CoresPerSocket=8\n CPUAlloc=0 CPUTot=32 CPULoad=0.06\n AvailableFeatures=(null)\n ActiveFeatures=(null)\n Gres=(null)\n NodeAddr=med0740 NodeHostName=med0740 Version=20.02.0\n OS=Linux 3.10.0-1062.12.1.el7.x86_64 #1 SMP Tue Feb 4 23:02:59 UTC 2020\n RealMemory=1 AllocMem=0 FreeMem=174388 Sockets=2 Boards=1\n State=IDLE ThreadsPerCore=2 TmpDisk=0 Weight=1 Owner=N/A MCS_label=N/A\n Partitions=debug\n BootTime=2020-03-05T00:54:15 SlurmdStartTime=2020-03-05T16:23:25\n CfgTRES=cpu=32,mem=1M,billing=32\n AllocTRES=\n CapWatts=n/a\n CurrentWatts=0 AveWatts=0\n ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s\nqueue states
$ squeue\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n$ squeue -u holtgrem_c\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\nnode resources
$ sinfo -o \"%20N %10c %10m %25f %10G \"\nadditional resources such as GPUs
$ sinfo -o \"%N %G\"\nlisting job details
$ scontrol show job 225\nJobId=225 JobName=bash\n UserId=XXX(135001) GroupId=XXX(30069) MCS_label=N/A\n Priority=4294901580 Nice=0 Account=(null) QOS=normal\n JobState=FAILED Reason=NonZeroExitCode Dependency=(null)\n Requeue=1 Restarts=0 BatchFlag=0 Reboot=0 ExitCode=130:0\n RunTime=00:16:27 TimeLimit=14-00:00:00 TimeMin=N/A\n SubmitTime=2020-03-23T11:34:26 EligibleTime=2020-03-23T11:34:26\n AccrueTime=Unknown\n StartTime=2020-03-23T11:34:26 EndTime=2020-03-23T11:50:53 Deadline=N/A\n SuspendTime=None SecsPreSuspend=0 LastSchedEval=2020-03-23T11:34:26\n Partition=gpu AllocNode:Sid=hpc-login-1:1918\n ReqNodeList=(null) ExcNodeList=(null)\n NodeList=med0301\n BatchHost=med0301\n NumNodes=1 NumCPUs=2 NumTasks=0 CPUs/Task=1 ReqB:S:C:T=0:0:*:*\n TRES=cpu=2,node=1,billing=2\n Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*\n MinCPUsNode=1 MinMemoryNode=0 MinTmpDiskNode=0\n Features=(null) DelayBoot=00:00:00\n OverSubscribe=OK Contiguous=0 Licenses=(null) Network=(null)\n Command=bash\n WorkDir=XXX\n Power=\n TresPerNode=gpu:tesla:4\n MailUser=(null) MailType=NONE\nhost:~$ squeue\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) \n 1177 medium bash jweiner_ R 4-21:52:24 1 med0127 \n 1192 medium bash jweiner_ R 4-07:08:40 1 med0127 \n 1209 highmem bash mkuhrin_ R 2-01:07:17 1 med0402 \n 1210 gpu bash hilberta R 1-10:30:34 1 med0304 \n 1213 long bash schubacm R 1-09:42:27 1 med0127 \n 2401 gpu bash ramkem_c R 1-05:14:53 1 med0303 \n 2431 medium ngs_mapp holtgrem R 1-05:01:41 1 med0127 \n 2437 critical snakejob holtgrem R 1-05:01:34 1 med0135 \n 2733 debug bash schubacm R 7:36:42 1 med0127 \n 3029 critical ngs_mapp holtgrem R 5:59:07 1 med0127 \n 3030 critical snakejob holtgrem R 5:56:23 1 med0134 \n 3031 critical snakejob holtgrem R 5:56:23 1 med0137 \n 3032 critical snakejob holtgrem R 5:56:23 1 med0137 \n 3033 critical snakejob holtgrem R 5:56:23 1 med0138 \n 3034 critical snakejob holtgrem R 5:56:23 1 med0138 \n 3035 critical snakejob holtgrem R 5:56:20 1 med0139 \n 3036 critical snakejob holtgrem R 5:56:20 1 med0139 \n 3037 critical snakejob holtgrem R 5:56:20 1 med0140 \n 3038 critical snakejob holtgrem R 5:56:20 1 med0140 \n 3039 critical snakejob holtgrem R 5:56:20 1 med0141 \n 3040 critical snakejob holtgrem R 5:56:20 1 med0141 \n 3041 critical snakejob holtgrem R 5:56:20 1 med0142 \n 3042 critical snakejob holtgrem R 5:56:20 1 med0142 \n 3043 critical snakejob holtgrem R 5:56:20 1 med0143 \n 3044 critical snakejob holtgrem R 5:56:20 1 med0143 \n 3063 long bash schubacm R 4:12:37 1 med0127 \n 3066 long bash schubacm R 4:11:47 1 med0127 \n 3113 medium ngs_mapp holtgrem R 1:52:33 1 med0708 \n 3118 medium snakejob holtgrem R 1:50:38 1 med0133 \n 3119 medium snakejob holtgrem R 1:50:38 1 med0703 \n 3126 medium snakejob holtgrem R 1:50:38 1 med0706 \n 3127 medium snakejob holtgrem R 1:50:38 1 med0144 \n 3128 medium snakejob holtgrem R 1:50:38 1 med0144 \n 3133 medium snakejob holtgrem R 1:50:35 1 med0147 \n 3134 medium snakejob holtgrem R 1:50:35 1 med0147 \n 3135 medium snakejob holtgrem R 1:50:35 1 med0148 \n 3136 medium snakejob holtgrem R 1:50:35 1 med0148 \n 3138 medium snakejob holtgrem R 1:50:35 1 med0104 \nhost:~$ squeue -o \"%.10i %9P %20j %10u %.2t %.10M %.6D %10R %b\"\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(R TRES_PER_NODE\n 1177 medium bash jweiner_m R 4-21:52:22 1 med0127 N/A\n 1192 medium bash jweiner_m R 4-07:08:38 1 med0127 N/A\n 1209 highmem bash mkuhrin_m R 2-01:07:15 1 med0402 N/A\n 1210 gpu bash hilberta_c R 1-10:30:32 1 med0304 gpu:tesla:4\n 1213 long bash schubacm_c R 1-09:42:25 1 med0127 N/A\n 2401 gpu bash ramkem_c R 1-05:14:51 1 med0303 gpu:tesla:1\n 2431 medium ngs_mapping holtgrem_c R 1-05:01:39 1 med0127 N/A\n 2437 critical snakejob.ngs_mapping holtgrem_c R 1-05:01:32 1 med0135 N/A\n 2733 debug bash schubacm_c R 7:36:40 1 med0127 N/A\n 3029 critical ngs_mapping holtgrem_c R 5:59:05 1 med0127 N/A\n 3030 critical snakejob.ngs_mapping holtgrem_c R 5:56:21 1 med0134 N/A\n 3031 critical snakejob.ngs_mapping holtgrem_c R 5:56:21 1 med0137 N/A\n 3032 critical snakejob.ngs_mapping holtgrem_c R 5:56:21 1 med0137 N/A\n 3033 critical snakejob.ngs_mapping holtgrem_c R 5:56:21 1 med0138 N/A\n 3034 critical snakejob.ngs_mapping holtgrem_c R 5:56:21 1 med0138 N/A\n 3035 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0139 N/A\n 3036 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0139 N/A\n 3037 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0140 N/A\n 3038 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0140 N/A\n 3039 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0141 N/A\n 3040 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0141 N/A\n 3041 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0142 N/A\n 3042 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0142 N/A\n 3043 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0143 N/A\n 3044 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0143 N/A\n 3063 long bash schubacm_c R 4:12:35 1 med0127 N/A\n 3066 long bash schubacm_c R 4:11:45 1 med0127 N/A\n 3113 medium ngs_mapping holtgrem_c R 1:52:31 1 med0708 N/A\n 3118 medium snakejob.ngs_mapping holtgrem_c R 1:50:36 1 med0133 N/A\n 3119 medium snakejob.ngs_mapping holtgrem_c R 1:50:36 1 med0703 N/A\n 3126 medium snakejob.ngs_mapping holtgrem_c R 1:50:36 1 med0706 N/A\n 3127 medium snakejob.ngs_mapping holtgrem_c R 1:50:36 1 med0144 N/A\n 3128 medium snakejob.ngs_mapping holtgrem_c R 1:50:36 1 med0144 N/A\n 3133 medium snakejob.ngs_mapping holtgrem_c R 1:50:33 1 med0147 N/A\n 3134 medium snakejob.ngs_mapping holtgrem_c R 1:50:33 1 med0147 N/A\n 3135 medium snakejob.ngs_mapping holtgrem_c R 1:50:33 1 med0148 N/A\n 3136 medium snakejob.ngs_mapping holtgrem_c R 1:50:33 1 med0148 N/A\n 3138 medium snakejob.ngs_mapping holtgrem_c R 1:50:33 1 med0104 N/A\nhost:~$ sinfo\nPARTITION AVAIL TIMELIMIT NODES STATE NODELIST \ndebug* up 8:00:00 11 drain med[0707,0709-0710,0740-0742,0744-0745,0749,0752,0755] \ndebug* up 8:00:00 8 mix med[0104,0127,0133-0135,0703,0706,0708] \ndebug* up 8:00:00 10 alloc med[0137-0144,0147-0148] \ndebug* up 8:00:00 103 idle med[0105-0124,0136,0145-0146,0151-0164,0201-0264,0704-0705] \nmedium up 7-00:00:00 11 drain med[0707,0709-0710,0740-0742,0744-0745,0749,0752,0755] \nmedium up 7-00:00:00 8 mix med[0104,0127,0133-0135,0703,0706,0708] \nmedium up 7-00:00:00 10 alloc med[0137-0144,0147-0148] \nmedium up 7-00:00:00 103 idle med[0105-0124,0136,0145-0146,0151-0164,0201-0264,0704-0705] \nlong up 28-00:00:0 11 drain med[0707,0709-0710,0740-0742,0744-0745,0749,0752,0755] \nlong up 28-00:00:0 8 mix med[0104,0127,0133-0135,0703,0706,0708] \nlong up 28-00:00:0 10 alloc med[0137-0144,0147-0148] \nlong up 28-00:00:0 103 idle med[0105-0124,0136,0145-0146,0151-0164,0201-0264,0704-0705] \ncritical up 7-00:00:00 11 drain med[0707,0709-0710,0740-0742,0744-0745,0749,0752,0755] \ncritical up 7-00:00:00 8 mix med[0104,0127,0133-0135,0703,0706,0708] \ncritical up 7-00:00:00 10 alloc med[0137-0144,0147-0148] \ncritical up 7-00:00:00 103 idle med[0105-0124,0136,0145-0146,0151-0164,0201-0264,0704-0705] \nhighmem up 14-00:00:0 1 mix med0402 \nhighmem up 14-00:00:0 3 idle med[0401,0403-0404] \ngpu up 14-00:00:0 2 mix med[0303-0304] \ngpu up 14-00:00:0 2 idle med[0301-0302] \nsacct","text":"Perform queries to the Slurm accounting information.
Representative Example
hpc-login-1:~$ sacct -j 1607103\n JobID JobName Partition Account AllocCPUS State ExitCode\n------------ ---------- ---------- ---------- ---------- ---------- --------\n1607103 wgs_sv_an+ medium 1 PENDING 0:0\nThe sacct command displays information from the Slurm accounting service. The Slurm scheduler only knows about active or completing (very recently active) jobs. The accouting system also knows about currently running jobs so it is the more robust way to query information about jobs. However, not all information is available to the accouting system, so scontrol show job and squeue provide more information about current and pending jbos.
Slurm Documentation: sacct
Please also see the official Slurm documentation on sacct.
"},{"location":"slurm/commands-sacct/#important-arguments","title":"Important Arguments","text":"Also see all important arguments of the sbatch command.
--jobs -- The job(s) to query for.--format -- Define attributes to retrieve.--long -- Get a lot of information from the database, consider to pipe into | less -S.sacct over scontrol and squeue.sattach","text":"The sattach command allows you to connect the standard input, output, and error streams to your current terminals ession.
Representative Example
hpc-login-1:~$ sattach 12345.0\n[...output of your job...]\nmed0211:~$ [Ctrl-C]\nhpc-login-1:~$\nPress Ctrl-C to detach from the current session. Please note that you will have to give the job ID as well as step step ID. For most cases, simply append \".0\" to your job ID.
Slurm Documentation: sattach
Please also see the official Slurm documentation on srun.
"},{"location":"slurm/commands-sattach/#important-arguments","title":"Important Arguments","text":"--pty -- Execute task zero in pseudo terminal.--verbose -- Increase verbosity of sattach.sbatch","text":"The sbatch command allows you to put a job into the scheduler's queue to be executed at a later time.
Representative Example
# Execute job.sh in partition medium with 4 threads and 4GB of RAM total for a\n# running time of up to one day.\nhpc-login-1:~$ sbatch --partition=medium --mem=4G --ntasks 4 --time=1-00 job.sh\nSubmitted batch job JOB_ID\nThe command will create a batch job and add it to the queue to be executed at a later point in time.
Slurm Documentation: sbatch
Please also see the official Slurm documentation on sbatch.
"},{"location":"slurm/commands-sbatch/#important-arguments","title":"Important Arguments","text":"--array -- Submit jobs as array jobs. Also see the section [#array-jobs] below.--nodes -- The number of nodes to allocate. This is only given here as an important argument as the maximum number of nodes allocatable to any partition but mpi is set to one (1). This is done as there are few users on the BIH HPC that actually use multi-node paralleilsm. Rather, most users will use multi-core parallelism and might forget to limit the number of nodes which causes inefficient allocation of resources.--cpus-per-task -- This corresponds to the number of CPU cores allocated to each task.--mem -- The memory to allocate for the job. As you can define minimal and maximal number of tasks/CPUs/cores, you could also specify --mem-per-cpu and get more flexible scheduling of your job.--gres -- Generic resource allocation. On the BIH HPC, this is only used for allocating GPUS, e.g., with --gres=gpu:tesla:2, a user could allocate two NVIDIA Tesla GPUs on the same host (use a40 instead of tesla for the A40 GPUs).--licenses -- On the BIH HPC, this is used for the allocation of MATLAB 2016b licenses only.--partition -- The partition to run in. Also see the Job Scheduler section.--time -- Specify the running time, see man sbatch or the official Slurm documentation on srun for supported formats. **Please note that the DRMA API only accepts the hours:minutes format.--dependency -- Specify dependencies on other jobs, e.g., using --dependency afterok:JOBID to only execute if the job with ID JOBID finished successfully or --dependency after:JOBID to wait for a job to finish regardless of its termination status.--constraint -- Require one or more features from your node. On the BIH HPC, the processor generation is defined as a feature on the nodes, e.g., haswell, or special networking such as infiniband. You can have a look at /etc/slurm/slurm.conf on all configured features.--output -- The path to the output log file (by default joining stdout and stderr, see the man page on --error on how to redirect stderr separately). A various number of placeholders is available, see the \"filename pattern\" section of man sbatch or the official Slurm documentation on srun.--mail-type=<type> -- Send out notifications by email when an event occurs. Use FAIL to get emails when your job fails. Also see the documentation of sbatch in the Slurm manual.--mail-user=<email> -- The email address to send to. Must end in @charite.de, @mdc-berlin.de, or @bih-charite.de.Ensure your --output directory exists!
In the case that the path to the log/output file does not exist, the job will just fail. scontrol show job ID will report JobState=FAILED Reason=NonZeroExitCode. Regrettably, no further information is displayed to you as the user. Always check that the path to the directories in StdErr and StdOut exists when checking scontrol show job ID.
--job-nameAlso see the section Slurm Job Scripts on how to embed the sbatch parameters in #SBATCH lines.
If you have many (say, more than 10) similar jobs (e.g., when performing a grid search), you can also use array jobs. However, you should also consider whether it would make sense to increase the time of your jobs, e.g, to be at least ~10min.
You can submit array jobs by specifying -a EXPR or --array EXPR where EXPR is a range or a list (of course, you can also add this as an #SBATCH header in your job script). For example:
hpc-login-1 ~# sbatch -a 1-3 grid_search.sh\nhpc-login-1 ~# sbatch -a 1,2,5-10 grid_search.sh\nThis will submit grid_search.sh with certain variables set:
SLURM_ARRAY_JOB_ID -- the ID of the first jobSLURM_ARRAY_TASK_ID -- the index of the job in the arraySLURM_ARRAY_TASK_COUNT -- number of submitted jobs in arraySLURM_ARRAY_TASK_MAX -- higehst job array index valueSLURM_ARRAY_TASK_MIN -- lowest job array index valueUsing array jobs has several advantages:
Also see Slurm documentation on job arrays.
For example, if you submit sbatch --array=1-3 grid_search.sh and slurm responsds with Submitted batch job 36 then the script will be run three times with the following prameters set:
SLURM_JOB_ID=36\nSLURM_ARRAY_JOB_ID=36\nSLURM_ARRAY_TASK_ID=1\nSLURM_ARRAY_TASK_COUNT=3\nSLURM_ARRAY_TASK_MAX=3\nSLURM_ARRAY_TASK_MIN=1\n\nSLURM_JOB_ID=37\nSLURM_ARRAY_JOB_ID=36\nSLURM_ARRAY_TASK_ID=2\nSLURM_ARRAY_TASK_COUNT=3\nSLURM_ARRAY_TASK_MAX=3\nSLURM_ARRAY_TASK_MIN=1\n\nSLURM_JOB_ID=38\nSLURM_ARRAY_JOB_ID=36\nSLURM_ARRAY_TASK_ID=3\nSLURM_ARRAY_TASK_COUNT=3\nSLURM_ARRAY_TASK_MAX=3\nSLURM_ARRAY_TASK_MIN=1\nsbatch are governed by resource allocations, in particular:sbatch jobs have a maximal running time set,sbatch jobs have a maximal memory and number of cores set, andscontrol show job JOBID.scancel","text":"Terminate a running Slurm job.
Representative Example
hpc-login-1:~$ scancel 1703828\nhpc-login-1:~$\nThis command allows to terminate one or more running jobs (of course, non-superusers can only terminate their own jobs).
Slurm Documentation: scancel
Please also see the official Slurm documentation on srun.
"},{"location":"slurm/commands-scontrol/","title":"Slurm Command:scontrol","text":"The scontrol allows to query detailed information from the scheduler and perform manipulation. Object manipulation is less important for normal users.
Representative Example
hpc-login-1:~$ scontrol show job 1607103\nJobId=1607103 JobName=wgs_sv_annotation\n UserId=holtgrem_c(100131) GroupId=hpc-ag-cubi(5272) MCS_label=N/A\n Priority=748 Nice=0 Account=(null) QOS=normal\n [...]\nhpc-login-1:~$ scontrol show node med02[01-32]\nNodeName=med0201 Arch=x86_64 CoresPerSocket=8\n CPUAlloc=0 CPUTot=32 CPULoad=0.01\n AvailableFeatures=ivybridge,infiniband\n ActiveFeatures=ivybridge,infiniband\n [...]\nhpc-login-1:~$ scontrol show partition medium\nPartitionName=medium\n AllowGroups=ALL AllowAccounts=ALL AllowQos=ALL\n AllocNodes=ALL Default=NO QoS=medium\n DefaultTime=NONE DisableRootJobs=NO ExclusiveUser=NO GraceTime=0 Hidden=NO\n [...]\nThis command allows to query all information for an object from Slurm, e.g., jobs, nodes, or partitions. The command also accepts ranges of jobs and hosts. It is most useful to get the information of one or a few objects from the scheduler.
Slurm Documentation: scontrol
Please also see the official Slurm documentation on scontrol.
"},{"location":"slurm/commands-scontrol/#important-sub-commands","title":"Important Sub commands","text":"scontrol show job -- Show details on jobs.scontrol show partition -- Show details on partitions.scontrol show node -- Show details on nodes.scontrol help -- Show help.scontrol -- Start an interactive scontrol shell / REPL (read-eval-print loop).scontrol can only work on jobs that are pending (in the queue), running, or in \"completing' state.sacct command.sinfo","text":"The sinfo command allows you to query the current cluster status.
Representative Example
hpc-login-1:~$ sinfo\nPARTITION AVAIL TIMELIMIT NODES STATE NODELIST\n[...]\nmedium up 7-00:00:00 10 drain* med[0101-0103,0125-0126,0128-0132]\nmedium up 7-00:00:00 1 down* med0243\nmedium up 7-00:00:00 31 mix med[0104,0106-0122,0124,0133,0232-0233,0237-0238,0241-0242,0244,0263-0264,0503,0506]\nmedium up 7-00:00:00 5 alloc med[0105,0123,0127,0239-0240]\nmedium up 7-00:00:00 193 idle med[0134-0164,0201-0231,0234-0236,0245-0262,0501-0502,0504-0505,0507-0516,0601-0632,0701-0764]\n[...]\nhpc-login-1:$ sinfo --summarize\nPARTITION AVAIL TIMELIMIT NODES(A/I/O/T) NODELIST\ndebug* up 8:00:00 38/191/11/240 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\nmedium up 7-00:00:00 38/191/11/240 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\nlong up 28-00:00:0 38/191/11/240 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\ncritical up 7-00:00:00 25/141/10/176 med[0101-0164,0501-0516,0601-0632,0701-0764]\nhighmem up 14-00:00:0 1/2/1/4 med[0401-0404]\ngpu up 14-00:00:0 3/0/1/4 med[0301-0304]\nmpi up 14-00:00:0 38/191/11/240 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\nThis command will summaries the state of nodes by different criteria (e.g., by partition or globally).
Slurm Documentation: sinfo
Please also see the official Slurm documentation on srun.
"},{"location":"slurm/commands-sinfo/#important-arguments","title":"Important Arguments","text":"Also see all important arguments of the sinfo command.
--summarize -- Summarize the node state by partition.--nodes -- Select the nodes to show the status for, e.g., display the status of all GPU nodes with sinfo -n med030[1-4].The most important node states are:
down -- node is marked as offlinedraining -- node will not accept any more jobs but has jobs running on itdrained -- node will not accept any more jobs and has no jobs running on it, but is not offline yetidle -- node is ready to run jobsallocated -- node is fully allocated (e.g., CPU, RAM, or GPU limit has been reached)mixed -- node is running jobs but there is space for moresqueue","text":"The squeue command allows you to view currently running and pending jobs.
Representative Example
hpc-login-1:~$ squeue\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 1583165 highmem 20200702 usr PD 0:00 1 (DependencyNeverSatisfied)\n 1605901 critical variant_ holtgrem PD 0:00 1 (DependencyNeverSatisfied)\n 1605902 critical variant_ holtgrem PD 0:00 1 (Dependency)\n 1605905 critical variant_ holtgrem PD 0:00 1 (DependencyNeverSatisfied)\n 1605916 critical wgs_sv_c holtgrem PD 0:00 1 (Dependency)\n 1607103 medium wgs_sv_a holtgrem PD 0:00 1 (DependencyNeverSatisfied)\n[...]\nSlurm Documentation: squeue
Please also see the official Slurm documentation on squeue.
"},{"location":"slurm/commands-squeue/#important-arguments","title":"Important Arguments","text":"--nodelist -- Only display jobs running on certain nodes (e.g., GPU nodes).--format -- Define the format to print, see man squeue for details. See below for a format string that includes the jobid, partition, job name, user name, job status, running time, number of nodes, number of CPU cores, and allocated GPUs.The following aliases in ~/.bashrc will allow you to print a long and informative squeue output with sq, pipe it into less with sql, get only your jobs (adjust the alias to your account) using sqme and pipe that into less with sqmel.
alias sq='squeue -o \"%.10i %9P %60j %10u %.2t %.10M %.6D %.4C %10R %b\" \"$@\"'\nalias sql='sq \"$@\" | less -S'\nalias sqme='sq -u YOURUSER_c_or_m \"$@\"'\nalias sqmel='sqme \"$@\" | less -S'\nsrun","text":"The srun command allows you to run a command now.
Representative Example
hpc-login-1:~$ srun --pty bash -i\nmed0201:~$\nThe command will perform a resource allocation with the scheduler (and wait until it has allocated the requested resources) first. Most importantly, you can specify the --pty argument which will connect the current terminal's standard output, error, and input to your current one. This allows you to run interactive jobs such as shells with srun --pty bash -i.
Slurm Documentation: srun
Please also see the official Slurm documentation on srun.
"},{"location":"slurm/commands-srun/#important-arguments","title":"Important Arguments","text":"Also see all important arguments of the sbatch command.
--pty -- Connect current terminal to the job's stdoud/stderr/stdin.--x11 -- Setup X11 forwarding.--immediate -- Immediately terminate if the resources to run the job are not available, do not wait.--test-only -- Don't run anything, but only estimate when the job would be scheduled.srun are governed by resource allocations, in particular:srun jobs have a maximal running time set,srun jobs have a maximal memory and number of cores set, andscontrol show job JOBID.In the sections Slurm Quickstart and Slurm Cheat Sheet, we have seen that sinfo and squeue allow for the compact display partitions/nodes and node information. In contrast, scontrol show job <id> and scontrol show partition <id> and scontrol show node <id> show comprehensive information that quickly gets hard to comprehend for multiple entries.
Now you might ask: is there anything in between? And: yes, there is.
You can tune the output of sinfo and squeue using parameters, in particular by providing format strings. All of this is described in the man pages of the commands that you can display with man sinfo and man squeue on the cluster.
sinfo Output","text":"Notable arguments of sinfo are:
-N, --Node -- uncompress the usual lines and display one line per node and partition.-s, --summarize -- compress the node state, more compact display.-R, --list-reasons -- for nodes that are not up, display reason string provided by admin.-o <fmt>, --format=<fmt> -- use format string for display.The most interesting argument is -o/--format. The man page lists the following values that are used when using other arguments. In other words, many of the display modifications could also be applied with -o/--format.
default \"%#P %.5a %.10l %.6D %.6t %N\"\n--summarize \"%#P %.5a %.10l %.16F %N\"\n--long \"%#P %.5a %.10l %.10s %.4r %.8h %.10g %.6D %.11T %N\"\n--Node \"%#N %.6D %#P %6t\"\n--long --Node \"%#N %.6D %#P %.11T %.4c %.8z %.6m %.8d %.6w %.8f %20E\"\n--list-reasons \"%20E %9u %19H %N\"\n--long --list-reasons\n \"%20E %12U %19H %6t %N\"\nThe best way to learn more about this is to play around with sinfo -o, starting out with one of the format strings above. Details about the format strings are described in man sinfo. Some remarks here:
%<num><char> displays the value represented by <char> padded with spaces to the right such that a width of <num> is reached,%.<num><char> displays the value represented by <char> padded with spaces to the left such that a width of <num> is reached, and%#<char> displays the value represented by <char> padded with spaces to the max length of the value represented by <char> (this is a \"virtual\" value, used internally only, you cannot use this and you will have to place an integer here).For example, to create a grouped display with reasons for being down use:
hpc-login-1:~$ sinfo -o \"%10P %.5a %.10l %.16F %40N %E\"\nPARTITION AVAIL TIMELIMIT NODES(A/I/O/T) NODELIST REASON\ndebug* up 8:00:00 0/0/16/16 med[0703-0710,0740-0742,0744-0745,0749,0 bogus node\ndebug* up 8:00:00 18/98/0/116 med[0104-0124,0127,0133-0148,0151-0164,0 none\nmedium up 7-00:00:00 0/0/16/16 med[0703-0710,0740-0742,0744-0745,0749,0 bogus node\nmedium up 7-00:00:00 18/98/0/116 med[0104-0124,0127,0133-0148,0151-0164,0 none\nlong up 28-00:00:0 0/0/16/16 med[0703-0710,0740-0742,0744-0745,0749,0 bogus node\nlong up 28-00:00:0 18/98/0/116 med[0104-0124,0127,0133-0148,0151-0164,0 none\ncritical up 7-00:00:00 0/0/16/16 med[0703-0710,0740-0742,0744-0745,0749,0 bogus node\ncritical up 7-00:00:00 18/98/0/116 med[0104-0124,0127,0133-0148,0151-0164,0 none\nhighmem up 14-00:00:0 0/4/0/4 med[0401-0404] none\ngpu up 14-00:00:0 3/1/0/4 med[0301-0304] none\nsqueue Output","text":"The standard squeue output might yield the following
hpc-login-1:~$ squeue | head\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 3149 medium variant_ holtgrem PD 0:00 1 (Dependency)\n 1177 medium bash jweiner_ R 6-03:32:41 1 med0127\n 1192 medium bash jweiner_ R 5-12:48:57 1 med0127\n 1210 gpu bash hilberta R 2-16:10:51 1 med0304\n 1213 long bash schubacm R 2-15:22:44 1 med0127\n 2401 gpu bash ramkem_c R 2-10:55:10 1 med0303\n 3063 long bash schubacm R 1-09:52:54 1 med0127\n 3066 long bash schubacm R 1-09:52:04 1 med0127\n 3147 medium ngs_mapp holtgrem R 1-03:13:42 1 med0148\nLooking at man squeue, we learn that the default format strings are:
default \"%.18i %.9P %.8j %.8u %.2t %.10M %.6D %R\"\n-l, --long \"%.18i %.9P %.8j %.8u %.8T %.10M %.9l %.6D %R\"\n-s, --steps \"%.15i %.8j %.9P %.8u %.9M %N\"\nThis looks a bit wasteful. Let's cut down on the padding of the job ID and expand on the job name and remove some right paddings.
hpc-login-1:~$ squeue -o \"%.6i %9P %30j %.10u %.2t %.10M %.6D %R %b\" | head\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 3149 medium variant_calling holtgrem_c PD 0:00 1 (Dependency)\n 1177 medium bash jweiner_m R 6-03:35:55 1 med0127\n 1192 medium bash jweiner_m R 5-12:52:11 1 med0127\n 1210 gpu bash hilberta_c R 2-16:14:05 1 med0304\n 1213 long bash schubacm_c R 2-15:25:58 1 med0127\n 2401 gpu bash ramkem_c R 2-10:58:24 1 med0303\n 3063 long bash schubacm_c R 1-09:56:08 1 med0127\n 3066 long bash schubacm_c R 1-09:55:18 1 med0127\n 3147 medium ngs_mapping holtgrem_c R 1-03:16:56 1 med0148\nNow display how many of our internal projects still exist.
hpc-login-1:~$ squeue -o \"%.6i %9P %30j %.10u %.2t %.10M %.6D %10R %s\" | head\nThe next steps are (TODO):
This page describes how to create SLURM job scripts.
SLURM job scripts look as follows. On the top you have lines starting with #SBATCH. These appear as comments to bash scripts. These lines are interpreted by sbatch in the same way as command line arguments. That is, when later submitting the script with sbatch my-job.sh you can either have the parameter to the sbatch call or in the file.
Multi-Node Allocation in Slurm
Classically, jobs on HPC systems are written in a way that they can run on multiple nodes at once, using the network to communicate. Slurm comes from this world and when allocating more than one CPU/core, it might allocate them on different nodes. Please use --nodes=1 to force Slurm to allocate them on a single node.
Creating the Script
host:example$ cat >my-job.sh <<\"EOF\"\n#!/bin/bash\n#\n#SBATCH --job-name=this-is-my-job\n#SBATCH --output=output.txt\n#\n#SBATCH --ntasks=1\n#SBATCH --nodes=1\n#SBATCH --time=10:00\n#SBATCH --mem-per-cpu=100M\n\ndate\n\nhostname\n>&2 echo \"Hello World\"\n\nsleep 1m\n\ndate\nEOF\nAlso see the SLURM Rosetta Stone for more options.
Submit, Look at Queue & Result
host:example$ sbatch script.sh \nSubmitted batch job 315\nhost:example$ squeue -u holtgrem_c\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) \n 315 debug this-is- holtgrem R 0:40 1 med0127 \nhost:example$ sleep 2m\nhost:example$ squeue -u holtgrem_c\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) \nhost:example$ cat output.txt \nWed Mar 25 13:30:56 CET 2020\nmed0127\nHello World\nWed Mar 25 13:31:56 CET 2020\nMemory allocation is one of the topics that users find confusing most often. This section first gives some technical background and then explains how to implement this properly with Slurm on the BIH HPC.
"},{"location":"slurm/memory-allocation/#technical-background","title":"Technical Background","text":"Technical Background Summary
Main memory used to be one of the most important topics when programming, as computers had so little. There is the infamous quote \"640KB ought ot be enough for anybody\" wrongly attribute to Bill Gates which refers to the fact that early computers could only address that amount of memory. In MS DOS, one had to use special libraries for a program to use more memory. Today, computers are very fast and memory is plentiful and people can (rightfully) forget about memory allocation ... as long as they don't use \"much\" memory by today's standards.
The Linux operating system differentiates between the following types of memory:
Note that above we are talking about processes, not Slurm jobs yet. Let us look at this in detail:
Each program uses some kind of memory management. For example, in C the malloc and free functions manually allocate and free memory while in Java, R, and Python, memory allocation and release is done automatically using a concept called garbage collection. Each program starts with a certain virtual memory size, that is the amount of memory it can address, say 128MB. When the program allocates memory, the memory allocation mechanism will check whether it has sufficient space left. If not, it will request an increase in virtual memory from the operating system, e.g., to 256MB. If this fails then the program can try to handle the error, e.g., terminate gracefully, but many programs will just panic and stop. Otherwise, the program will get access to more memory and happily continue to run.
However, programs can allocate humonguous amounts of virtual memory and only use a little. Memory is organized in \"pages\" (classically these are 4096 bytes each, but can be larger using so-called \"huge page\" features). The operating system tracks which memory pages are actually used by a process. The total size of these pages is called the resident set size: the amount of memory that is actually currently used by a program. Programs can also mark pages as unused again, thus freeing resident memory and can also decrease their virtual memory.
In some cases it is still interesting to use swap memory. Here, the contents of resident memory are copied to disk by the operating system. This process is completely transparent to the program; the data remains available at the original positions in the virtual memory! However, accessing it will take some time as it must be read back into main memory from the disk. In this way, it was possible for a computer with 4MB of RAM and a disk of 100MB to run programs that used 8MB. Of course, this was only really useable for programs that ran in the background. One could really feel the latency if a graphical program was using swapped memory (you could actually hear the hard drive working). Today, swap storage is normally only relevant when put your computer into hibernation. Given the large main memory on the cluster nodes, their small local hard drives (just used for loading the operating system), and the extreme slowness involved in using swapped memory, the BIH HPC nodes have no swap memory allocated.
Most HPC users will also use shared memory, at least implicitly. Whenever a program uses fork to create a subprocess (BTW, this is not a thread), the program can chose to \"copy\" its current address space. The second process then has access to the same memory than the parent process in a copy-on-write fashion. This allows, for example, pre-loading a database, and also allows the use of already loaded library code by the child process as well. If the child process writes to the copy-on-write memory of the parent, the relevant memory page will be copied and attributed to the child.
Two or more processes can share the same memory explicitly. This is usually used for inter-process communication but the Bowtie program uses it for sharing the memory of indices. For example, the Python multiprocessing module will use this, including if you have two MPI processes running on the same host.
Memory is also separated into segments, the most interesting ones are heap and stack memory. For compiled languages, memory can be allocated on either. For C, an int variable will be allocated on the stack. Every time you call a function, a stack frame is created in memory to hold the local variables and other information for the duration of the function execution. The stack thus grows through function calls made by your program and shrinks when the functions return. The stack size for a process is limited (by ulimit -s) and a program that goes too deep (e.g., via infinite recursion) will be terminated by the operating system if it exceeds this limit. Again in C, int * ptr = (int *)malloc(10 * sizeof(int)); will allocate memory for one variable (an integer pointer) on the stack and memory for 10 integers on the heap. When the function returns, the ptr variable on the stack will be freed but to free the array of integers, you'd have to call free(ptr). If the memory is not freed then this constitutes a memory leak, but that is another topic.
Other relevant segments are code, where the compiled code lives, and data, where static data such as strings displayed to the user are stored. As a side node, in interpreted languages such as R or Python, the code and data segments will refer to the code and data of Python while the actual program text will be on the heap.
"},{"location":"slurm/memory-allocation/#interlude-memory-in-java","title":"Interlude: Memory in Java","text":"Memory in Java Summary
-XX:MaxHeapSize=<size> (e.g., <size>=2G) for your program and only tune the other parameters if neededJava's memory management provides for some interesting artifacts. When running simple Java programs, you will never run into this but if you need to use gigabytes of memory in Java then you will have to learn a bit about Java memory management. This is the case when running GATK programs, for example.
As different operating systems handle memory management differently, the Java virtual machine does its own memory management to provide a consistent interface. The following three settings are important in governing memory usage of Java:
-Xmx<size>/-XX:MaxHeapSize=<size> -- the maximal Java heap size-Xms<size>/-XX:InitialHeapSize=<size> -- the initial Java heap size-Xss<size>/-XX:ThreadStackSize=<size> -- maximal stack size available to a Java thread (e.g., the main thread)Above, <size> is a memory specification, either in bytes or with a suffix, e.g., 80M, or 1G.
On startup, Java does roughly the following:
Memory freed by the Java garbage collector can be re-used by other Java objects (rss remains the same) or be freed in the operating system (rss decreases). The Java VM program itself will also consume memory on the OS stack but that is negligible.
Overall, the Java VM needs to store in main memory:
In the BIH HPC context, the following is recommended to:
Memory Allocation in Slurm Summary
--mem=<size> (e.g., <size>=3G) to allocate memory per nodesrun and batch sbatch jobs are governed by Slurm memory allocationOur Slurm configuration uses Linux cgroups to enforce a maximum amount of resident memory. You simply specify it using --mem=<size> in your srun and sbatch command.
In the (rare) case that you provide more flexible number of threads (Slurm tasks) or GPUs, you could also look into --mem-per-cpu and --mem-per-gpu. The official Slurm sbatch manual is quite helpful, as is man sbatch on the cluster command line.
Slurm (or rather Linux via cgroups) will track all memory started by all jobs by your process. If each process works independently (e.g., you put the output through a pipe prog1 | prog2) then the amount of memory consumed will at any given time be the sum of the RSS of both processes at that time. If your program uses fork, which uses memory in a copy-on-write fashion, the shared memory is of course only counted once. Note that Python's multiprocessing does not use copy on write: its data will be explicitly copied and consume additional memory. Refer to the Scipy/Numpy/Pandas etc. documentation on how to achieve parallelism without copying too much data.
The amount of virtual memory that your program can reserve is only \"virtually\" unlimited (pun not intended). However, in practice, the operating system will not like you allocating more than physically available. If your program attempts to allocate more memory than requested via Slurm, your program will be killed.
This is reported to you in the Slurm job output log as something like:
slurmstepd: error: Detected 1 oom-kill event(s) in step <JOB ID>.batch cgroup. Some of your processes may have been killed by the cgroup out-of-memory handler.\nYou can inspect the amount of memory available on each node in total with sinfo --format \"%.10P %.10l %.6D %.6m %N\", as shown below.
$ sinfo --format \"%.10P %.10l %.6D %.6m %N\"\n PARTITION TIMELIMIT NODES MEMORY NODELIST\n debug* 8:00:00 240 128722 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\n medium 7-00:00:00 240 128722 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\n long 28-00:00:0 240 128722 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\n critical 7-00:00:00 176 128722 med[0101-0164,0501-0516,0601-0632,0701-0764]\n highmem 14-00:00:0 4 515762 med[0401-0404]\n gpu 14-00:00:0 4 385215 med[0301-0304]\n mpi 14-00:00:0 240 128722 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\nMemory Accounting in Slurm Summary
sacct -j JOBID --format=JobID,MaxRSS to display the RSS usage of your programsacct -j JOBID --format=Elapsed,AllocCPUs,TotalCPU to display information about CPU usageWhile Slurm runs your job, it collects information about the job such as the running time, exit status, and memory usage. This information is available through the scheduling system via the squeue and scontrol commands, but only while the job is pending execution, executing, or currently completing. After job completion, the information is only available through the Slurm accounting system.
You can query information about jobs, e.g., using sacct:
$ sacct -j 1607166\n JobID JobName Partition Account AllocCPUS State ExitCode\n------------ ---------- ---------- ---------- ---------- ---------- --------\n1607166 snakejob.+ critical 16 COMPLETED 0:0\n1607166.bat+ batch 16 COMPLETED 0:0\n1607166.ext+ extern 16 COMPLETED 0:0\nThis shows that the job with ID 1607166 with a job ID starting with snakejob. has been run in the critical partition, been allocated 16 cores and had an exit code of 0:0. For technical reasons, there is a batch and an extern sub step. Actually, Slurm makes it possible to run various steps in one batch as documented in the Slurm documentation.
The sacct command has various command-line options that you can read about via man sacct or in the Slurm documentation. We can use --brief/-b to show only a brief summary.
$ sacct -j 1607166 --brief\n JobID State ExitCode\n------------ ---------- --------\n1607166 COMPLETED 0:0\n1607166.bat+ COMPLETED 0:0\n1607166.ext+ COMPLETED 0:0\nSimilarly, you can use --long to display extended information (see the manual for the displayed columns). Very long report lines can be piped into less -S for easier display. You can fine-tune the information to display with a format string to --format:
$ sacct -j 1607166 --format=JobID,ReqMem,MaxRSS,Elapsed,TotalCPU,AllocCPUS\n JobID ReqMem MaxRSS Elapsed TotalCPU AllocCPUS\n------------ --------- ---------- ---------- ---------- ----------\n1607166 60Gn 13:07:31 7-16:21:29 16\n1607166.bat+ 60Gn 4314560K 13:07:31 7-16:21:29 16\n1607166.ext+ 60Gn 0 13:07:31 00:00.001 16\nFrom this command, we can read that we allocate 60GB memory of memory per node (suffix n, here Gn for gigabytes per node) and the maximum RSS is reported as 4.3GB. You can use this information to fine-tune your memory allocations. As a side-remark, a suffic c indicates the memory per core (e.g., that could be60Gc)
Further, the program ran for 13 hours and 7 minutes with allocated 16 CPU cores and consumed a total of 7 days, 16 hours, and 21 minutes of CPU time. Thus, a total of 10,061 CPU minutes were spent in 787 minutes wall-clock time. This yields an overall empirical degree of parallelism of about 10061 / 787 = 14, and a parallel efficiency of 14 / 16 = 88%. The discussion of parallel efficiency is a topic not covered here.
However, you can use the awk script below to compute the empirical parallelism (EmpPar) and the parallel efficiency (ParEff). The script also displays the difference I requested, and used RSS (DiffRSS). The script can be found here.
$ sacct -j 1607166 --format=JobID,ReqMem,MaxRSS,Elapsed,TotalCPU,AllocCPUS \\\n | awk -f quick-sacct.awk\n JobID ReqMem MaxRSS Elapsed TotalCPU AllocCPUS EmpPar ParEff DiffMEM\n------------ ---------- ---------- ---------- ---------- ---------- --------- -------- --------\n1607166 60Gn 13:07:31 7-16:21:29 16 0.00 0.00 -\n1607166.bat+ 60Gn 4314560K 13:07:31 7-16:21:29 16 14.05 0.88 55.89\n1607166.ext+ 60Gn 0 13:07:31 00:00.001 16 0.00 0.00 -\nThe BIH HPC uses the Slurm scheduling system for resource allocation. This section of the manual attempts to give an overview of what scheduling is and how to use the Slurm scheduler. For more detailed information, you will have to refer to the Slurm website and the Slurm man pages (e.g., by entering man sbatch or man srun on the HPC terminal's command line).
For a quick introduction and hands-on examples, please see the manual sections
Also, make sure that you are aware of our How-To: Debug Software and How-To: Debug Software on HPC Systems guides in the case that something goes wrong.
"},{"location":"slurm/overview/#annotated-contents","title":"Annotated Contents","text":"srun -- running parallel jobs nowsbatch -- submission of batch jobsscancel -- stop/kill jobssinfo -- display information about the Slurm clustersqueue -- information about pending and running jbosscontrol -- detailed information (and control)sacct -- access Slurm accounting information (pending, running, and past jobs)Many other facilities run Slurm clusters and make their documentation available on the internet. We list some that we found useful below. However, be aware that Slurm is a highly configurable and extensible system. Other sites may have different configurations and plugins enabled than we have (or might even have written custom plugins that are not available at BIH). In any case, it's always useful to look \"\u00fcber den Tellerrand\".
man Pages - web versions of Unix manual (man) pages.Create an interactive bash session (srun will run bash in real-time, --pty connects its stdout and stderr to your current session).
hpc-login-1:~$ srun --pty bash -i\nmed0740:~$ echo \"Hello World\"\nHello World\nmed0740:~$ exit\nhpc-login-1:~$\nNote you probably want to longer running time for your interactive jobs. This way, your jobs can run for up to 28 days. This will make your job be routed automatically into the long partition as it is the only one that can fit your job.
hpc-login-1:~$ srun --pty --time 28-00 bash -i\nmed0740:~$\nPro-Tip: Using Bash aliases for quick access.
hpc-login-1:~$ alias slogin=\"srun --pty bash -i\"\nhpc-login-1:~$ slogin\nmed0740:~$ exit\nhpc-login-1:~$ cat >>~/.bashrc <<\"EOF\"\n# Useful aliases for logging in via Slurm\nalias slogin=\"srun --pty bash -i\"\nalias slogin-x11=\"srun --pty --x11 bash -i\"\nEOF\nCreate an interactive R session on the cluster (assuming conda is active and the environment my-r is created, e.g., with conda create -n my-r r).
hpc-login-1:~$ conda activate my-r\nhpc-login-1:~$ srun --pty R\nR version 3.6.2 (2019-12-12) -- \"Dark and Stormy Night\"\nCopyright (C) 2019 The R Foundation for Statistical Computing\n[...]\nType 'demo()' for some demos, 'help()' for on-line help, or\n'help.start()' for an HTML browser interface to help.\nType 'q()' to quit R.\n\n\n> Sys.info()[\"nodename\"]\n nodename\n\"med0740\"\n> q()\nSave workspace image? [y/n/c]:\nhpc-login-1:~$\nCreate an interactive iPython session on the cluster (assuming conda is active and the environment my-python is created, e.g., with conda create -n my-python python=3 ipython).
hpc-login-1:~$ conda activate my-python\nhpc-login-1:~$ srun --pty ipython\nPython 3.8.2 | packaged by conda-forge | (default, Mar 5 2020, 17:11:00)\nType 'copyright', 'credits' or 'license' for more information\nIPython 7.13.0 -- An enhanced Interactive Python. Type '?' for help.\n\nIn [1]: import socket; socket.gethostname()\nOut[1]: 'med0740'\n\nIn [2]: exit\nhpc-login-1:~$\nAllocate 4 cores (default is 1 core), and a total of 4GB of RAM on one node (alternatively use --mem-per-cpu to set RAM per CPU); sbatch accepts the same argument.
hpc-login-1:~$ srun --cpus-per-task=4 --nodes=1 --mem=4G --pty bash\nmed0740:~$ export | grep SLURM_CPUS_ON_NODE\n4\nmed0740:~$ your-parallel-script --threads 4\nSubmit an R script to the cluster in batch mode (sbatch schedules the job for later execution).
hpc-login-1:~$ cat >job-script.sh <<\"EOF\"\n#!/bin/bash\necho \"Hello, I'm running on $(hostname) and it's $(date)\"\nEOF\nhpc-login-1:~$ sbatch job-script.sh\nSubmitted batch job 7\n\n# Some time later:\nhpc-login-1:~$ cat slurm-7.out\nHello, I'm running on med0740 and it's Fri Mar 6 07:36:42 CET 2020\nhpc-login-1:~$\nHint
Read this in particular if you want to know why your job does not get scheduled and you see Reason=ReqNodeNotAvail,_Reserved_for_maintenance in scontrol show job .
Administration registers maintenances with the Slurm scheduler as so-called reservations. You can see the current reservations with scontrol show reservation. The following is a scheduled reservation affecting ALL nodes of the cluster.
# scontrol show reservation\nReservationName=root_13 StartTime=2021-09-07T00:00:00 EndTime=2021-09-09T00:00:00 Duration=2-00:00:00\n Nodes=hpc-cpu-[1-36],med[0101-0116,0201-0264,0301-0304,0401-0404,0501-0516,0601-0632,0701-0764]\n NodeCnt=236 CoreCnt=5344 Features=(null) PartitionName=(null)\n Flags=MAINT,IGNORE_JOBS,SPEC_NODES,ALL_NODES TRES=cpu=10176\n Users=root Groups=(null) Accounts=(null) Licenses=(null) State=INACTIVE BurstBuffer=(null) Watts=n/a\n MaxStartDelay=(null)\nYou will also be notified when logging into the login nodes, e.g.,
--\n ***NOTE: 1 scheduled maintenance(s)***\n\n 1: 2021-09-07 00:00:00 to 2021-09-09 00:00:00 ALL nodes\n\nYou jobs do not start because of \"Reserved_for_maintenance\"?\nSlurm jobs will only start if they do not overlap with scheduled reservations.\nMore information:\n\n - https://bihealth.github.io/bih-cluster/slurm/reservations/\n - https://bihealth.github.io/bih-cluster/admin/maintenance/\n--\nMaintenance reservations will block the affected nodes (or even the whole cluster) for jobs. If there is a maintenance in one week then your job must have an end time before the reservation starts. By this, the job gives a guarantee to the scheduler that it will not interfer with the maintenance reservation.
For example, scontrol show job JOBID might report the following
JobId=4011580 JobName=snakejob\n UserId=USER(UID) GroupId=GROUP(GID) MCS_label=N/A\n Priority=1722 Nice=0 Account=GROUP QOS=normal\n JobState=PENDING Reason=ReqNodeNotAvail,_Reserved_for_maintenance Dependency=(null)\n Requeue=1 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0\n RunTime=00:00:00 TimeLimit=28-00:00:00 TimeMin=N/A\n SubmitTime=2021-08-30T09:01:01 EligibleTime=2021-08-30T09:01:01\n AccrueTime=2021-08-30T09:01:01\n StartTime=2021-09-09T00:00:00 EndTime=2021-10-07T00:00:00 Deadline=N/A\n SuspendTime=None SecsPreSuspend=0 LastSchedEval=2021-08-30T10:20:40\n Partition=long AllocNode:Sid=172.16.35.153:5453\n ReqNodeList=(null) ExcNodeList=(null)\n NodeList=(null)\n NumNodes=1-1 NumCPUs=8 NumTasks=8 CPUs/Task=1 ReqB:S:C:T=0:0:*:*\n TRES=cpu=8,mem=4G,node=1,billing=8\n Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*\n MinCPUsNode=1 MinMemoryNode=4G MinTmpDiskNode=0\n Features=(null) DelayBoot=00:00:00\n OverSubscribe=OK Contiguous=0 Licenses=(null) Network=(null)\n Power=\n NtasksPerTRES:0\nLook out for the Reason line:
Reason=ReqNodeNotAvail,_Reserved_for_maintenance\nThis job is scheduled to run up to 4 weeks and has been submitted on 2021-08-30.
Right now the following reservation is active
# scontrol show reservation\nReservationName=root_13 StartTime=2021-09-07T00:00:00 EndTime=2021-09-09T00:00:00 Duration=2-00:00:00\n Nodes=hpc-cpu-[1-36],med[0101-0116,0201-0264,0301-0304,0401-0404,0501-0516,0601-0632,0701-0764]\n NodeCnt=236 CoreCnt=5344 Features=(null) PartitionName=(null)\n Flags=MAINT,IGNORE_JOBS,SPEC_NODES,ALL_NODES TRES=cpu=10176\n Users=root Groups=(null) Accounts=(null) Licenses=(null) State=INACTIVE BurstBuffer=(null) Watts=n/a\n MaxStartDelay=(null)\nThus, the scheduler decided to set a StartTime of the job to 2021-09-09T00:00:00, which is the end time of the reservation. Effectively, the job is forced to run outside the maintenance reservation.
You can resolve this by using a --time= parameter to srun or sbatch such that the job ends before the maintenance reservation starts.
Rosetta Stone?
The Rosetta Stone is a stone slab that carries the same text in Egyptian hieroglyphs and ancient Greek. This was key for decyphering Egyptian hieroglyphs in the 18th century. Nowadays, the term is often used to label translation tables such as the one below.
The table below shows some SGE commands and their Slurm equivalents.
User Command SGE Slurm remote loginqrsh/qlogin srun --pty bash run interactively N/A srun --pty program submit job qsub script.sh sbatch script.sh delete job qdel job-id scancel job-id job status by job id N/A squeue --job job-id detailed job status qstat -u '*' -j job-id sstat job-id job status of your jobs qstat squeue --me job status by user qstat -u user squeue -u user hold job qhold job-id scontrol hold job-id release job qrls job-id scontrol release job-id queue list qconf -sql scontrol show partitions node list qhost sinfo -N OR scontrol show nodes cluster status qhost -q sinfo show node resources N/A sinfo \"%n %G\" Job Specification SGE Slurm script directive marker #$ #SBATCH (run in queue) -q queue -p queue allocated nodes N/A -N min[-max] allocate cores -pe smp count -n count limit running time -l h_rt=time -t days-hh:mm:s redirectd stdout -o file -o file redirect stderr -e file -e file combine stdout/stderr -j yes -o without -e copy environment -V --export=ALL\\|NONE\\|variables email notification -m abe --mail-type=events send email to -M email --mail-user=email job name -N name --job-name=name restart job -r yes|no --requeue|--no-requeue working directory -wd path --workdir run exclusively -l exclusive --exclusive OR --shared allocate memory -l h_vmem=size --mem=mem OR --mem-per-cpu=mem wait for job -hold_jid jid --depend state:job select target host -l hostname=host1\\|host1 --nodelist=nodes AND/OR --exclude allocate GPU -l gpu=1 --gres=gpu:tesla:count or --gres=gpu:a40:count"},{"location":"slurm/snakemake/","title":"Snakemake with Slurm","text":"This page describes how to use Snakemake with Slurm.
"},{"location":"slurm/snakemake/#prerequisites","title":"Prerequisites","text":"source miniforge/bin/activate.We first create a new environment snakemake-slurm and activate it. We need the snakemake package for this. For snakemake 8, we additionally need snakemake-executor-plugin-slurm.
host:~$ conda create -y -n snakemake-slurm 'snakemake>=8.24.1' snakemake-executor-plugin-slurm\n[...]\n#\n# To activate this environment, use\n#\n# $ conda activate snakemake-slurm\n#\n# To deactivate an active environment, use\n#\n# $ conda deactivate\nhost:~$ conda activate snakemake-slurm\n(snakemake-slurm) host:~$\nWe create a workflow and ensure that it works properly with multi-threaded Snakemake (no cluster submission here!)
host:~$ mkdir -p snake-slurm\nhost:~$ cd snake-slurm\nhost:snake-slurm$ cat >Snakefile <<\"EOF\"\nrule default:\n input: \"the-result.txt\"\n\nrule mkresult:\n output: \"the-result.txt\"\n shell: r\"sleep 2m; touch the-result.txt\"\nEOF\nhost:snake-slurm$ snakemake --cores=1\n[...]\nhost:snake-slurm$ ls\nSnakefile the-result.txt\nhost:snake-slurm$ rm the-result.txt\nSimply use snakemake --profile=cubi-v1 and the Snakemake resource configuration as shown below. This works for both snakemake version 7 and 8.
(For version 7, you can also use snakemake --cluster='sbatch ...' command instead, but this is discouraged.)
Note that we sneaked in a sleep 2m? In a second terminal session, we can see that the job has been submitted to SLURM indeed.
host:~$ squeue -u holtgrem_c\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 325 debug snakejob holtgrem R 0:47 1 med0127\nThe cubi-v1 profile (stored in /etc/xdg/snakemake/cubi-v1 on all cluster nodes) supports the following specification in your Snakemake rule:
threads: the number of threads to execute the job onresources.mem/resources.mem_mb: the memory to allocate for the whole job, OR resources.mem_per_thread: the memory to allocate for each thread.resources.time: the running time of the rule, in a syntax supported by Slurm, e.g. HH:MM:SS or D-HH:MM:SSresources.partition: the partition to submit your job into (Slurm will pick a fitting partition for you by default)resources.nodes: the number of nodes to schedule your job on (defaults to 1 and you will want to keep that value unless you want to use MPI)You will need Snakemake >=7.0.2 for this.
For more information on resources, see snakemake resources and slurm resources.
Here is how to call Snakemake:
# snakemake --profile=cubi-v1 --jobs 1\n--sdm conda (snakemake 8) or --use-conda (snakemake 7). To set rule-specific resources:
rule myrule:\n threads: 1\n resources:\n mem='8G',\n time='04:00:00',\n input: # ...\n output: # ...\n shell: # ...\nYou can combine this with Snakemake resource callables, of course:
def myrule_mem(wildcards, attempt):\n mem = 2 * attempt\n return '%dG' % mem\n\nrule snps:\n threads: 1\n resources:\n mem=myrule_mem,\n time='04:00:00',\n input: # ...\n output: # ...\n shell: # ...\nBy default, slurm will write log files into the working directory of snakemake, which will look like slurm-$jobid.out.
To change this behaviour, the environment variable SBATCH_DEFAULTS can be set to re-route the --output parameter. If you want to write your files into slurm_logs with a filename pattern of $name-$jobid for instance, consider the following snippet for your submission script:
#!/bin/bash\n#\n#SBATCH --job-name=snakemake_main_job\n#SBATCH --ntasks=1\n#SBATCH --nodes=1\n#SBATCH --time=48:10:00\n#SBATCH --mem-per-cpu=300M\n#SBATCH --output=slurm_logs/%x-%j.log\n\nmkdir -p slurm_logs\nexport SBATCH_DEFAULTS=\" --output=slurm_logs/%x-%j.log\"\n\ndate\nsrun snakemake --use-conda -j1 --profile=cubi-v1\ndate\nThe name of the snakemake slurm job will be snakemake_main_job, the name of the jobs spawned from it will be called after the rule name in the Snakefile.
This section describes how Slurm handles temporary files on the local disk.
Temporary Files Best Practices
See Best Practices: Temporary Files for information how to use temporary files effectively.
"},{"location":"slurm/temporary-files/#slurm-behaviour","title":"Slurm Behaviour","text":"Our Slurm configuration has the following behaviour.
"},{"location":"slurm/temporary-files/#environment-variable-tmpdir","title":"Environment Variable TMPDIR","text":"Slurm itself will by default not change the TMPDIR environment variable but retain the variable's value from the srun or sbatch call.
/tmp Directories","text":"The only place where users can write data to on local storage of the compute nodes is /tmp.
Storage is a consumable shared resource as the storage used by one job cannot use another job. It is thus critical that Slurm cleans up after each job such that all space on the local node is available to the next job. This is done using the job_container/tmpfs Slurm plugin.
This plugin creates a so-called Linux namespace for each job and creates a bind mount of /tmp to a location on the local storage. This mount is only visible to the currently running job and each job, even of the same user, get their own /tmp. After a job terminates, Slurm will remove the directory and all of its content.
There is a notable exception. If you use ssh to connect to a node rather than using srun or sbatch, you will see the system /tmp directory and can also write to it. This usage of storage is not tracked and consequently you can circumvent the Slurm quota management. Using /tmp in this fashion (i.e., outside of Slurm-controlled jobs) is prohibited. If it cannot be helped (e.g., if you need to run some debugging application that needs to create FIFO or socket files) then keep usage of /tmp outside of Slurm job below 100MB.
localtmp","text":"Enforcing localtmp Gres
From January 31, we will enforce the allocated storage in /tmp on the local disk with quotas. Jobs writing to /tmp beyond the quota in the job allocation will not function properly and probably crash with \"out of disk quota\" messages.
Slurm tracks the available local storage above 100MB on nodes in the localtmp generic resource (aka Gres). The resource is counted in steps of 1MB, such that a node with 350GB of local storage would look as follows in scontrol show node:
hpc-login-1 # scontrol show node hpc-cpu-1\nNodeName=hpc-cpu-1 Arch=x86_64 CoresPerSocket=24\n [...]\n Gres=localtmp:350K\n [...]\n CfgTRES=cpu=96,mem=360000M,billing=96,gres/localtmp=358400\n [...]\nEach job is automaticaly granted 100MB of storage on the local disk which is sufficient for most standard programs. If your job needs more temporary storage then you should either
$HOME/scratch volume (see Best Practices: Temporary Files)localtmp generic resource (described here)You can allocate the resource with --gres=localtmp:SIZE where SIZE is given in MB.
hpc-login-1 # srun --gres=localtmp:100k --pty bash -i\nhpc-cpu-1 # scontrol show node hpc-cpu-1\nNodeName=hpc-cpu-1 Arch=x86_64 CoresPerSocket=24\n [...]\n Gres=localtmp:250K\n [...]\n CfgTRES=cpu=96,mem=360000M,billing=96,gres/localtmp=358400\n [...]\n AllocTRES=cpu=92,mem=351G,gres/localtmp=102400\n [...]\nThe first output tells us about the resource configured to be available to user jobs and the last line show us that 100k=102400 MB of local storage are allocated.
You can also see the used resources in the details of your job:
scontrol show job 14848\nJobId=14848 JobName=example.sh\n [...]\n TresPerNode=gres:localtmp:100k\nMake sure to connect to the login node with X11 forwarding.
host:~$ ssh -X -l user_c hpc-login-1.cubi.bihealth.org\nOnce connected to the login node, pass the --x11 flag.
hpc-login-1:~$ srun --pty --x11 xterm\nWe set quite restrictive quotas for user homes, but in exchange you get file system snapshots and mirroring. Your home folder should therefore only be used for scripts, your user config, and other small files. Everything else should be stored in the work or scratch subdirectories, which effectively link to your group's shared storage space. This document describes some common pitfalls and how to circumvent them.
Hint
The tilde character (~) is shorthand for your home directory.
Various programs are used to depositing large folders in a user's home and can quickly use up your allotted storage quota. These include:
~/.local/lib/python*~/R/x86_64-pc-linux-gnu-library~/ondemandPlease note that directories whose name is starting with a dot are not shown by the normal ls command, but require the ls -a flag. You can search your home folder for large directories like so:
$ du -shc ~/.* ~/* --exclude=.. --exclude=.\nYou should move these locations to your work folder and create symbolic links in their place. Conda installations should be installed in work from the very beginning as they do not react well to being moved around.
Here is an example for the .local folder.
$ mv ~/.local ~/work/.local\n$ ln -s ~/work/.local ~/.local\nAnother usual culprit is the hidden .cache directory which contains temporary files. This folder can be moved to the scratch volume in a similar manner as described above.
$ mv ~/.cache ~/scratch/.cache\n$ ln -s ~/scratch/.cache ~/.cache\nImportant
Files placed in your scratch directory will be automatically removed after 2 weeks. Do not place any valuable files in there.
Please use hpc-transfer-1 and hpc-transfer-2 for moving large amounts of files. This not only leaves the compute notes available for actual computation, but also has no risk of your jobs being killed by Slurm. You should also use tmux to not risk connection loss during long running transfers.
Define source and target location and copy contents. Please replace the parts in curly brackets with your actual folder names. It is important to end paths with a trailing slash (/) as this is interpreted by sync as \u201call files in this folder\u201d.
$ SOURCE=/data/gpfs-1/work/projects/{my_project}/\n$ TARGET=/data/cephfs-2/unmirrored/projects/{my-project}/\n$ rsync -ahP --stats --dry-run $SOURCE $TARGET\nRemove the --dry-run flag to start the actual copying process.
Important
File ownership information will be lost during this process. This is due to non-root users not being allowed to change ownership of arbitrary files. If this is a problem for you, please contact our admins again after completing this step.
Perform a second rsync to check if all files were successfully transferred. Paranoid users might want to add the --checksum flag to rsync or use hashdeep. Please note the flag --remove-source-files which will do exactly as the name suggests, but leaves empty directories behind.
$ rsync -ahX --stats --remove-source-files --dry-run $SOURCE $TARGET\n--dry-run flag to start the actual deletion.$ find $SOURCE -type f | wc -l\n0\n$ rm -r $SOURCE\nWarning
When defining your SOURCE location, do not use the * wildcard character. It will not match hidden (dot) files and leave them behind. Its better to use a trailing slash which matches \u201cAll files in this folder\u201d.
All files within your own work directory can be transferred as follows. Please replace parts in curly braces with your cluster user name.
$ SOURCE=/data/gpfs-1/work/users/{username}/\n$ TARGET=/data/cephfs-1/home/users/{username}/work/\n$ rsync -ahP --stats --dry-run $SOURCE $TARGET\nNote
The --dry-run flag lets you check that rsync is working as expected without copying any files. Remove it to start the actual transfer.
Perform a second rsync to check if all files were successfully transferred. Paranoid users might want to add the --checksums flag or use hashdeep. Please note the flag --remove-source-files which will do exactly as the name suggests, but leaves empty directories behind.
$ rsync -ahP --stats --remove-source-files --dry-run $SOURCE $TARGET\n$ find $SOURCE -type f | wc -l\n0\nConda installations tend not to react well to moving their main folder from its original location. There are numerous ways around this problem which are described here.
A simple solution we can recommend is this:
Install a fresh version of conda or mamba in your new work folder. Don't forget to first remove the conda init block in ~/.bashrc.
$ nano ~/.bashrc\n$ conda init\n$ conda config --set auto_activate_base false\nYou can then use your new conda to export your old environments by specifying a full path like so:
$ conda env export -p /fast/work/user/$USER/miniconda/envs/<env_name> -f <env_name>.yaml\n--no-builds flag. Finally re-create your old environments from the yaml files:
$ conda env create -f {environment.yml}\nOutdated
This document is only valid for the old, third-generation file system and will be removed soon. Quotas of our new CephFS storage are communicated via the HPC Access web portal.
As described elsewhere, all data in your user, group, and project volumes is subject to quotas. This page quickly shows how to query for the current usage of data volume and file counts for your user, group, and projects.
"},{"location":"storage/querying-storage/#query-for-user-data-and-file-usage","title":"Query for User Data and File Usage","text":"The file /etc/bashrc.gpfs-quota contains some Bash functions that you can use for querying the quota usage. This file is automatically sourced in all of your Bash sessions.
For querying your user's data and file usage, enter the following command:
# bih-gpfs-quota-user holtgrem_c\nYou will get a report as follows. As soon as usage reaches 90%, the data/file usage will be highlighted in yellow. If you pass 99%, the data/file usage will be highlighted in red.
=================================\nQuota Report for: user holtgrem_c\n=================================\n\n DATA quota GR- FILES quota GR-\nENTITY NAME FSET USED SOFT HARD ACE USED SOFT HARD ACE\n------- ---------- ------- ----- ---- ----- ----- --- ----- ---- ----- ----- ---\nusers holtgrem_c home 103M 10% 1.0G 1.5G - 2.5k 25% 10k 12k -\nusers holtgrem_c work 639G 62% 1.0T 1.1T - 1.0M 52% 2.0M 2.2M -\nusers holtgrem_c scratch 42G 0% 200T 220T - 207k 0.1% 200M 220M -\n[...]\n# bih-gpfs-report-quota group ag_someag\n=================================\nQuota Report for: group ag_someag\n=================================\n\n DATA quota GR- FILES quota GR-\nENTITY NAME FSET USED SOFT HARD ACE USED SOFT HARD ACE\n------- ---------- ------- ----- ---- ----- ----- --- ----- ---- ----- ----- ---\ngroups ag_someag home 0 0% 1.0G 1.5G - 4 0% 10k 12k -\ngroups ag_someag work 349G 34% 1.0T 1.5T - 302 0% 2.0M 2.2M -\ngroups ag_someag scratch 0 0% 200T 220T - 1 0% 200M 220M -\n\n[...]\n# bih-gpfs-report-quota project someproj\n==================================\nQuota Report for: project someproj\n==================================\n\n DATA quota GR- FILES quota GR-\nENTITY NAME FSET USED SOFT HARD ACE USED SOFT HARD ACE\n------- ---------- ------- ----- ---- ----- ----- --- ----- ---- ----- ----- ---\ngroups someproj home 0 0% 1.0G 1.5G - 4 0% 10k 12k -\ngroups someproj work 349G 34% 1.0T 1.5T - 302 0% 2.0M 2.2M -\ngroups someproj scratch 0 0% 200T 220T - 1 0% 200M 220M -\n\n[...]\nThe scratch space is automatically cleaned up nightly with the following mechanism.
scratch folder are created and retained for 3 days.Warning
We specifically use the mtime attribute to determine if files in scratch should be cleaned up. Copying or downloading files to scratch while preserving the original mtime might lead to unexpected results.
This document describes the forth iteration of the file system structure on the BIH HPC cluster. It was made necessary because the previous file system was no longer supported by the manufacturer and we since switched to distributed Ceph storage.
Important
For now, the old, third-generation file system is still mounted at /fast. It will be decommissioned soon, please consult this document describing the migration process!
There are the following three entities on the cluster:
Each user, group, and project can have storage folders in different locations.
"},{"location":"storage/storage-locations/#data-types-and-storage-tiers","title":"Data Types and Storage Tiers","text":"Files stored on the HPC fall into one of three categories:
Home folders store programs, scripts, and user config i.\u00a0e. long-lived and very important files. Loss of this data requires to redo manual work (like programming).
Work folders store data of potentially large size which has a medium life time and is important. Examples are raw sequencing data and intermediate results that are to be kept (e.\u00a0g. sorted and indexed BAM files). Work data requires time-consuming actions to be restored, such as downloading large amounts of data or long-running computation.
Scratch folder store temporary files with a short life-time. Examples are temporary files (e.\u00a0g. unsorted BAM files). Scratch data is created to be removed eventually.
Ceph storage comes in two types which differ in their I/O speed, total capacity, and cost. They are called Tier 1 and Tier 2 and sometimes hot storage and warm storage. In the HPC filesystem they are mounted in /data/cephfs-1 and /data/cephfs-2.
Storage quotas are imposed in these locations to restrict the maximum size of folders. Amount and utilization of quotas is communicated via the HPC Access web portal.
"},{"location":"storage/storage-locations/#home-directories","title":"Home Directories","text":"Location: /data/cephfs-1/home/
Only users have home directories on Tier 1 storage. This is the starting point when starting a new shell or SSH session. Important config files are stored here as well as analysis scripts and small user files. Home folders have a strict storage quota of 1\u00a0GB.
"},{"location":"storage/storage-locations/#work-directories","title":"Work Directories","text":"Location: /data/cephfs-1/work/
Groups and projects have work directories on Tier 1 storage. User home folders contain a symlink to their respective group's work folder. Files shared within a group/project are stored here as long as they are in active use. Work folders are generally limited to 1\u00a0TB per group. Project work folders are allocated on an individual basis.
"},{"location":"storage/storage-locations/#scratch-space","title":"Scratch Space","text":"Location: /data/cephfs-1/scratch/
Groups and projects have scratch space on Tier 1 storage. User home folders contain a symlink to their respective group's scratch space. Meant for temporary, potentially large data e.\u00a0g. intermediate unsorted or unmasked BAM files, data downloaded from the internet etc. Scratch space is generally limited to 10\u00a0TB per group. Projects are allocated scratch on an individual basis. Files in scratch will be automatically removed 2 weeks after their creation.
"},{"location":"storage/storage-locations/#tier-2-storage","title":"Tier 2 Storage","text":"Location: /data/cephfs-2/
This is where big files go when they are not in active use. Groups are allocated 10 TB of Tier 2 storage by default. File quotas here can be significantly larger as space is much cheaper and more abundant than on Tier 1.
Note
Tier 2 storage is currently not accessible from HPC login nodes.
"},{"location":"storage/storage-locations/#overview","title":"Overview","text":"Tier Function Path Default Quota 1 User home/data/cephfs-1/home/users/<user> 1 GB 1 Group work /data/cephfs-1/work/groups/<group> 1 TB 1 Group scratch /data/cephfs-1/scratch/groups/<group> 10 TB 1 Project work /data/cephfs-1/work/projects/<project> On request 1 Project scratch /data/cephfs-1/scratch/projects/<project> On request 2 Group /data/cephfs-2/unmirrored/groups/<group> 10 TB 2 Project /data/cephfs-2/unmirrored/projects/<project> On request 2 Group /data/cephfs-2/mirrored/groups/<group> On request 2 Project /data/cephfs-2/mirrored/projects/<project> On request"},{"location":"storage/storage-locations/#snapshots-and-mirroring","title":"Snapshots and Mirroring","text":"Snapshots are incremental copies of the state of the data at a particular point in time. They provide safety against various \"Ops, did I just delete that?\" scenarios, meaning they can be used to recover lost or damaged files. Depending on the location and Tier, CephFS creates snapshots in different frequencies and retention plans.
Location Path Retention policy Mirrored User homes/data/cephfs-1/home/users/ Hourly for 48 h, daily for 14 d yes Group/project work /data/cephfs-1/work/ Four times a day, daily for 5 d no Group/project scratch /data/cephfs-1/scratch/ Daily for 3 d no Group/project mirrored /data/cephfs-2/mirrored/ Daily for 30 d, weekly for 16 w yes Group/project unmirrored /data/cephfs-2/unmirrored/ Daily for 30 d, weekly for 16 w no Some parts of Tier 1 and Tier 2 snapshots are also mirrored into a separate fire compartment within the data center. This provides an additional layer of security i.\u00a0e. physical damage to the servers.
"},{"location":"storage/storage-locations/#accessing-snapshots","title":"Accessing Snapshots","text":"To access snapshots simply navigate to the .snap/ sub-folder of the respective location. This special folder exists on all levels of the CephFS file hierarchy, so even in your user home directory. Inside you will find one folder per snapshot created and in those a complete replica of the respective folder at the time of snapshot creation.
For example:
/data/cephfs-1/home/.snap/<some_snapshot>/users/<your_user>/ same as:/data/cephfs-1/home/users/<your_user>/.snap/<some_snapshot>/data/cephfs-1/work/.snap/<some_snapshot>/groups/<your_group>//data/cephfs-2/unmirrored/.snap/<some_snapshot>/projects/<your_project>/Here is a simple example of how to restore a file:
$ cd /data/cephfs-2/unmirrored/groups/cubi/.snap/scheduled-2024-03-11-00_00_00_UTC/\n$ ls -l\nimportant_file.txt\n$ cp important_file.txt /data/cephfs-2/unmirrored/groups/cubi/\n/data/cephfs-1Important
We will remove access to /fast on most cluster nodes following September 30th.
Files on the cluster's main storage /data/gpfs-1 aka. /fast will move to a new file system. That includes users' home directories, work directories, and work-group directories. Once files have been moved to their new locations, /fast will be retired.
Simultaneously we will move towards a more unified naming scheme for project and group folder names. From now on, all such folders names shall be in kebab-case. This is Berlin after all. Group folders will also be renamed, removing the \"ag_\" prefix.
Detailed communication about the move will be communicated via the cluster mailinglist and the user forum. For technical help, please consult the Data Migration Tips and tricks.
"},{"location":"storage/storage-migration/#why-is-this-happening","title":"Why is this happening?","text":"/fast is based on a high performance proprietary hardware (DDN) & file system (GPFS). The company selling it has terminated support which also means buying replacement parts will become increasingly difficult.
There are two file systems set up to replace /fast, named Tier 1 and Tier 2 after their difference in I/O speed:
/fast ever was, but it only has about 75\u00a0% of its usable capacity.The Hot storage Tier 1 is reserved for files requiring frequent random access, user homes, and scratch. Tier 2 (Warm storage) should be used for everything else. Both file systems are based on the open-source, software-defined Ceph storage platform and differ in the type of drives used. Tier 1 or Cephfs-1 uses NVME SSDs and is optimized for performance, Tier 2 or Cephfs-2 used traditional hard drives and is optimized for cost.
So these are the three terminologies in use right now:
/data/cephfs-1/data/cephfs-2More information about CephFS can be found here.
"},{"location":"storage/storage-migration/#new-file-locations","title":"New file locations","text":"Naturally, paths are going to change after files move to their new location. Due to the increase in storage quality options, there will be some more folders to consider.
"},{"location":"storage/storage-migration/#users","title":"Users","text":"/data/cephfs-1/home/users/<user>/data/cephfs-1/work/groups/<doe>/users/<user>/data/cephfs-1/scratch/groups/<doe>/users/<user>Important
User work & scratch spaces are now part of the user's group folder. This means, groups need to coordinate internally to distribute their allotted quota according to each user's needs.
The implementation is done via symlinks created by default when the user account is moved to its new destination:
~/work -> /data/cephfs-1/work/groups/<group>/users/<user>~/scratch -> /data/cephfs-1/scratch/groups/<group>/users/<user>/data/cephfs-1/work/groups/<group>/data/cephfs-1/scratch/groups/<group>/data/cephfs-2/unmirrored/groups/<group>/data/cephfs-1/work/projects/<project>/data/cephfs-1/scratch/projects/<project>Space on Tier 1 is limited. Your colleagues, other cluster users, and admins will be very grateful if you use it only for files you actively need to perform read/write operations on. This means main project storage should probably always be on Tier 2 with workflows to stage subsets of data onto Tier 1 for analysis.
These examples are based on our experience of processing diverse NGS datasets. Your mileage may vary but there is a basic principle that remains true for all projects.
"},{"location":"storage/storage-migration/#dna-sequencing-wes-wgs","title":"DNA sequencing (WES, WGS)","text":"Typical Whole Genome Sequencing data of a human sample at 100x coverage requires about 150 GB of storage, Whole Exome Sequencing files occupy between 6 and 30 GB. These large files require considerable I/O resources for processing, in particular for the mapping step. A prudent workflow for these kind of analysis would therefore be the following:
fastqs) from the Tier 2 location to Tier 1. seqtk is your friend!fastq files from Tier 2 to Tier 1. Run the your scripts on the whole dataset, and copy the results (bam or cram files) back to Tier 2.Tip
Don't forget to use your scratch area for transient operations, for example to sort your bam file after mapping. More information on how to efficiently set up your temporary directory here.
Analysis of RNA expression datasets are typically a long and iterative process, where the data must remain accessible for a significant period. However, there is usually no need to keep raw data files and mapping results available once the gene & transcripts counts have been generated. The count files are much smaller than the raw data or the mapped data, so they can live longer on Tier 1.
A typical workflow would be:
fastq files from Tier 2 to Tier 1.salmon or STAR, and store the results on Tier 2.R, using tximport and DESeq2 or featureCounts & edgeR, for example.R objects) and the output of salmon, STAR, or any mapper/aligner of your choice to Tier 2.Tip
If using STAR, don't forget to use your scratch area for transient operations. More information on how to efficiently set up your temporary directory here
The analysis workflow of bulk RNA & single cell dataset is conceptually similar: Large raw files need to be processed once and only the outcome of the processing (gene counts matrices) are required for downstream analysis. Therefore, a typical workflow would be:
fastq files from Tier 2 to Tier 1.Cell Ranger or alevin-fry, perform count matrix QC and store the results on Tier 2.seurat, scanpy, or Loupe Browser.There is no obvious workflow that covers most used cases for machine learning. However,
/fast to CephFS","text":"Best practices and tools will be provided.
"}]} \ No newline at end of file +{"config":{"lang":["en"],"separator":"[\\s\\-]+","pipeline":["stopWordFilter"]},"docs":[{"location":"","title":"Home","text":"Welcome to the user documentation of the BIH high-performance computing (HPC) cluster, also called HPC 4 Research. The BIH HPC cluster is managed by CUBI (Core Unit Bioinformatics). This documentation is maintained by BIH CUBI and the user community. It is a living document that you can update and add to. See How-To: Contribute to this Document for details.
The global table of contents is on the left, the one of the current page is on the right.
Additional resources
Read the following set of pages (in order) to learn how to get access and connect to the cluster.
Acknowledging BIH HPC Usage
Acknowledge usage of the cluster in your manuscript as \"Computation has been performed on the HPC for Research/Clinic cluster of the Berlin Institute of Health\". Please add your publications using the cluster to this list.
"},{"location":"#news-maintenance-announcements","title":"News & Maintenance Announcements","text":"hpc-mem-5 with 4 TB of RAM./fast on all non-transfer nodes.See Maintenance for a detailed list of current, planned, and previous maintenance and update work.
"},{"location":"#technical-details","title":"Technical Details","text":"If you are interested in how this HPC cluster is set up on a technical level, we got you covered. There is an entire section on this.
"},{"location":"#documentation-structure","title":"Documentation Structure","text":"The documentation is structured as follows:
Access to the BIH HPC cluster is conceptually based on user groups (also known as labs or units) and projects. Users have a relatively limited storage quota within their private home folder and store big data primarily within their group's work space or in project folders. Projects are collaborative efforts involving multiple PIs/groups and are allocated separate storage space on the cluster.
Independent group leaders at BIH/Charit\u00e9/MDC can request a group on the cluster and name group members. The work group leader (the group PI) bears the responsibility for the group's members and ensures that cluster policies and etiquette are followed. In brief: Fair usage rules apply and the cluster ist not to be abused for unethical or illegal purposes. Major and/or continued violations may lead to exclusion of the entire group.
The group leader may also name one delegate (typically an IT-savvy Post-Doc) who is thereby allowed to take decision about cluster usage and work group management on behalf of the group leader. The above mentioned responsibilities stay with the group leader.
Note
All cluster users are member of exactly one primary work group. This affiliation is usually defined by real life organisational structures within Charit\u00e9/BIH/MDC. Leaders of independent research groups (PIs) can apply for a new cluster work group as follows:
Important
Changes to an existing group (adding new users, changes in resources, etc.) can only be requested by group leaders and delegates.
"},{"location":"admin/getting-access/#form-new-group","title":"Form: New Group","text":"Example values are given in curly braces.
# Group \"ag-{doe}\"\nGroup leader/PI: {John Doe}\nDelegate [optional]: {Max Mustermann}\nPurpose of cluster usage [short]: {RNA-seq analysis in colorectal cancer}\n\nRequired resources:\n- Tier 1 storage: {1 TB}\n- Tier 1 scratch: {10 TB}\n- Tier 2 storage: {10 TB}\n\n# Users\n## User 1\n- first name: {John}\n- last name: {Doe}\n- affiliation: {Charit\u00e9, Department of Oncology}\n- institute email: {john.doe@charite.de}\n- user has account with\n - [ ] BIH\n - [x] Charite\n - [ ] MDC\n- BIH/Charit\u00e9/MDC user name: {doej}\n\n## User 2\n[etc.]\nExample values are given in curly braces.
# New user of AG {Doe}\n- first name: {Mia}\n- last name: {Smith}\n- affiliation: {Charit\u00e9, Department of Oncology}\n- institute email: {mia.smith@charite.de}\n- user has account with\n - [ ] BIH\n - [x] Charite\n - [ ] MDC\n- BIH/Charit\u00e9/MDC user name: {smithm}\nNotes
Projects are secondary user groups to enable:
Project creation can be initiated by group leaders and group delegates as follows:
Important
Changes to an existing project (adding new users, changes in resources, etc.) can only be requested by project owners and delegates. Please send us cluster user names for adding new project members.
"},{"location":"admin/getting-access/#form","title":"Form","text":"Example values are given in curly braces.
# Project \"{doe-dbgap-rna}\"\nProject owner: {John Doe}, {doej_c}\nDelegate [optional]: {Max Mustermann}, {musterm_c}\nPurpose of cluster usage [short]: {RNA-seq data from dbGAP}\n\nRequired resources:\n- Tier 1 work: {0 TB}\n- Tier 1 scratch: {0 TB}\n- Tier 2 storage: {1 TB}\n\nAdditional members (cluster user names):\n- {sorgls_c}\n- ...\nNotes
This page documents the current and known upcoming maintenance windows.
"},{"location":"admin/maintenance/#login-node-maintenance-december-9th-12th","title":"Login node maintenance, December 9th & 12th","text":"We will perform OS upgrades on both login nodes during this week. To keep disruption of service as minimal as possible upgrades will be staggered:
hpc-login-2 will be offline on Monday 9thhpc-login-1 will be offline on Thurday 12thPlease plan accordingly.
"},{"location":"admin/maintenance/#login-compute-maintenance-october-22nd-2024","title":"Login & Compute maintenance, October 22nd, 2024","text":"Kernel and Slurm upgrades. Keep an eye on our forum for updates.
"},{"location":"admin/maintenance/#login-compute-and-storage-maintenance-december-13-14-2022","title":"Login, Compute and Storage Maintenance, December 13-14, 2022","text":"All informationand updates regarding maintenance will be circulated on our forum https://hpc-talk.cubi.bihealth.org/c/announcements/5.
"},{"location":"admin/maintenance/#login-compute-and-storage-maintenance-march-22-23-2022","title":"Login, Compute and Storage Maintenance, March 22-23, 2022","text":"All COMPUTE nodes and STORAGE resources won't be reachable!
All nodes will be running in RESERVATION mode. This means you are still able to schedule new jobs on these nodes if their potential/allowed runtime does not extend into the maintenance window (Tuesday and Wednesday, March 22 and 23, all-day). For example, if you submit a job that can run up to 7 days after March 15 then the job will remain in \"pending/PD\" state giving the explanation of \"all nodes being reserved or unavailable\".
Issues of today's maintenance:
/tmp on login nodescephfs-2 switches (Tier 2 storage, not relevant for most users)IMPORTANT
Progress Thread on hpc-talk
"},{"location":"admin/maintenance/#drmaa-deprecation-march-2-2022","title":"DRMAA Deprecation, March 2, 2022","text":"scontrol show job JOBID and sacct -j JOBID.snakemake --profile=cubi-v1 instead of snakemake --drmaa \"...\".rule myrule:\n # ...\n threads: 8\n resources:\n time=\"12:00:00\",\n memory=\"8G\",\n # ...\nSchedulerParameters+=bf_max_job_user=50: backfill scheduler only considers 50 jobs of each user. This mitigates an issue with some users having too many jobs and thus other users' jobs don't get ahead in the queueEnforcePartLimits=ALL: jobs that don't fit into their partition are rejectedDependencyParameters=kill_invalid_depend: jobs that have dependencies set that cannot be fulfilled will be killedlocaltmp Resource, January 31, 2022","text":"localtmp resource for local storage above 100MB./tmp using Linux namespaces/cgroups. This greatly improves the reliability of cleaning up after jobs. (Technically, this is implemented using the Slurm job_container/tmpfs) plugin.Gres) \"localtmp\". In the future this will become a requirement. Also see Slurm: Temporary Files.hpc-login-1.cubi.bihealth.orghpc-login-2.cubi.bihealth.orghpc-portal.cubi.bihealth.orghpc-transfer-1.cubi.bihealth.orghpc-transfer-2.cubi.bihealth.orghpc-gpu-{5..7}. 28.08.5.The GPFS storage system has been upgraded to the latest version to make compatible with Enterprise Linux version 8.
"},{"location":"admin/maintenance/#slurm-upgrade-to-21080-september-8-2021","title":"Slurm upgrade to21.08.0, September 8, 2021","text":"Slurm has been upgraded to version 21.08.0.
All servers/nodes won't be reachable!
All nodes will be running in reservation mode. This means you are still able to schedule new jobs on these nodes if their potential/allowed runtime does not extend into the maintenance window (Tuesday and Wednesday, September 7 and 8, all-day). For example, if you submit a job that can run up to 7 days after August 30 then the job will remain in \"pending/PD\" state giving the explanation of \"all nodes being reserved or unavailable\".
If you already have a job running on any nodes that goes beyond September 7, 12:00 am (00:00 Uhr), this job will die.
"},{"location":"admin/maintenance/#renaming-of-gpu-high-memory-machines-scheduler-changes-september-7-2021","title":"Renaming of GPU & High Memory Machines & Scheduler Changes, September 7, 2021","text":"The GPU machines med030[1-4] have been renamed to hpc-gpu-[1-4]. The high memory machines med040[1-4] have been renamed to hpc-mem-[1-4]. It will probably take us some time to update all places in the documentation.
Further, the long partition has been changed to allow jobs with a maximum running time of 14 days.
staging partition, August 31, 2021","text":"We have installed 36 new nodes (in BETA mode) in the cluster called hpc-node-[1-36]. They have 48 cores (thus 96 hardware threads) each and have 360GiB of main memory available (for the hardware nerds, it's Intel(R) Xeon(R) Gold 6240R CPUs at 2.40GHz, featuring the cascadelake architecture).
Right now, they are only available in the staging partition. After some testing we will move them to the other partitions. We'd like to ask you to test them as well and report any issues to hpc-helpdesk@bih-charite.de. The nodes have been setup identically to the existing med0xxx nodes. We do not expect big changes but the nodes might not be as stable as other oness.
Here is how you can reach them.
hpc-login-1 # srun --immediate=5 --pty --time=24:00:00 --partition=staging bash -i\n[...]\nhpc-cpu-1 #\nNote that I'm specifying a maximal running time of 24h so the scheduler will end the job after 24 hours which is before the upcoming maintenance reservation begins. By default, the scheduler allocates 28 days to the job which means that the job cannot end before the reservation and will be scheduled to start after it. See Reservations / Maintenances for more information about maintenance reservations.
"},{"location":"admin/maintenance/#reservation-maintenance-display-on-login-august-30-2021","title":"Reservation / Maintenance Display on Login, August 30, 2021","text":"User will now be notified on login about maintenance, for example:
NOTE: scheduled maintenance(s)\n\n 1: 2021-09-07 00:00:00 to 2021-09-09 00:00:00 ALL nodes\n\nSlurm jobs will only start if they do not overlap with scheduled reservations.\nMore information:\n\n - https://bihealth.github.io/bih-cluster/slurm/reservations/\n - https://bihealth.github.io/bih-cluster/admin/maintenance/\nThe srun command will now behave as if --immediate=60 has been specified by default. It explains how to override this behaviour and possible reasons for job scheduling to fail within 60 seconds (reservations and full cluster).
We upgrade from 20.11.2 to 20.11.8 which contains some fixes for bugs that our users actually stumbled over. The change should be non-intrusive as it's only a patch-level update.
Following servers won't be reachable:
These nodes are running in reservation mode now. This means you are still able to schedule new jobs on these nodes if their potential/allowed runtime does not extend into the maintenance window (Tuesday, August 3, all-day). For example, if you submit a job that can run up to 7 days after July 26 then the job will remain in \"pending/PD\" state giving the explanation of \"all nodes being reserved or unavailable\". If you have a job running on any of the before mentioned nodes that goes beyond August 3, 12:00 am (00:00 Uhr), this job will die. We do not expect the remaining nodes to be affected. However, there remains a minor risk of unexpected downtime of other nodes.
"},{"location":"admin/maintenance/#server-reorganization-july-13-2021","title":"Server reorganization, July 13, 2021","text":"Affected servers are:
If you have a job running on any of the before mentioned nodes that goes beyond June 22, 6am, this job will die. We put a so-called Slurm reservation for the maintenance period. Any job that is scheduled before the maintenance and whose end time (start time + max running time) is not before the start of the maintenance will not be scheduled with the message ReqNodeNotAvail, Reserved for maintenance.
Affected servers are:
HPC 4 Research
Note
This task is currently being planned. No schedule has been fixed yet.
Note
This task is currently being planned. No schedule has been fixed yet.
/fast that currently points to /data/gpfs-1 on HPC 4 Research./data instead of /fast everywhere, e.g., /data/users/$NAME etc.Time: 6am-12am
/fast file system will be re-mounted to /data/gpfs-1./fast becomes a symbolic link to /data on all of the cluster.hpc-login-1.cubi.bihealth.org and login-2... instead of hpc-login-{1,2}.hpc-transfer-{1,2} which will be replaced by transfer-1.research.hpc.bihealth.org and transfer-2....med010[1-3] and med012[5-6].On June 3, we need to perform a network maintenance at 8 am.
If everything goes well, there might be a short delay in network packages and connections will survive. In this case, the maintenance will end 8:30 am.
Otherwise, the maintenance will finish by noon.
"},{"location":"admin/maintenance/#cluster-maintenance-with-downtime-june-16","title":"Cluster Maintenance with Downtime: June 16","text":"We need to schedule a full cluster downtime on June 16.
"},{"location":"admin/maintenance/#slurm-migration","title":"Slurm Migration","text":"We will switch to the Slurm workload scheduler (from the legacy SGE). The main reason is that Slurm allows for better scheduling of GPUs (and has loads of improvements over SGE), but the syntax is a bit different. Currently, our documentation is in an transient state. We are currently extending our Slurm-specific documentation.
SSH Key Management has switched to using Charite and MDC ActiveDirectory servers. You need to upload all keys by the end of April 2020.
Schedule
Feb 4, 2020: Keys are now also taken from central MDC/Charite servers. You do not need to contact us any more to update your keys (we cannot accelerate the process at MDC).May 1, 2020: Keys are now only taken from central MDC/Charite servers. You must upload your keys to central servers by then.Affected systems:
hpc-transfer-1hpc-transfer-2hpc-login-2The compute nodes are non-critical as we are taking them out of the queues now.
"},{"location":"admin/maintenance/#centos-76-upgrade-january-29-february-5","title":"CentOS 7.6 Upgrade, January 29, February 5","text":"Starting monday 03.09.2018 we will be performing rolling update of the cluster from CentOS 7.4 to CentOS 7.5. Since update will be performed in small bunches of nodes, the only impact you should notice is smaller number of nodes available for computation.
Also, for around two weeks, you can expect that your jobs can hit both CentOS 7.4 & CentOS 7.5 nodes. This should not impact you in any way, but if you encounter any unexpected behavior of the cluster during this time, please let us know.
At some point we will have to update the transfer, and login nodes. We will do this also in parts, so the you can switch to the other machine.
Key dates are:
18.09.2018 - hpc-login-1 & hpc-transfer-1 will not be available, and you should switch to hpc-login-2 & hpc-transfer-2 respectively.
25.09.2018 - hpc-login-2 & hpc-transfer-2 will not be available, and you should switch to hpc-login-1 & hpc-transfer-1 respectively.
Please also be informed that non-invasive maintenance this weekend which we announced has been canceled, so cluster will operate normally.
In case of any concerns, issues, do not hesitate to contact us via hpc-admin@bih-charite.de, or hpc-helpdesk@bih-charite.de.
"},{"location":"admin/maintenance/#june-18-2018-0600-1500","title":"June 18, 2018, 0600-1500","text":"Due to tasks we need to perform on BIH cluster, we have planned maintenance:
During maintenance we will perform several actions:
During maintenance whole cluster will not be usable, this includes:
Maintenance window is quite long, since we are dependent on external vendor. However, we will recover services as soon as possible.
We will keep you posted during maintenance with services status.
"},{"location":"admin/maintenance/#march-16-18-2018-mdc-it","title":"March 16-18, 2018 (MDC IT)","text":"MDC IT has a network maintenance from Friday, March 16 18:00 hours until Sunday March 18 18:00 hours.
This will affect connections to the cluster but no connections within the cluster.
"},{"location":"admin/maintenance/#january-17-2018-complete","title":"January 17, 2018 (Complete)","text":"STATUS: complete
The first aim of this window is to upgrade the cluster to CentOS 7.4 to patch against the Meltdown/Spectre vulnerabilities. For this, the login and transfer nodes have to be rebooted.
The second aim of this window is to reboot the file server to mitigate some NFS errors. For this, the SGE master has to be stopped for some time.
"},{"location":"admin/maintenance/#planprogress","title":"Plan/Progress","text":"(since January 2010)
This page describes strictly enforced policies valid on the BIH HPC clusters.
The aim of the HPC systems is to support the users in their scientific work and relies on their cooperation. First and foremost, the administration team enforces state of the art IT security and reliability practices through their organizational and operational processes and actions. We kindly ask user to follow the Cluster Etiquette describe below to allow for fair use and flexible access to the shared resources. Beyond this, policies are introduced or enforced only when required to ensure non-restrictive access to the resources themselves. Major or recurrent breaches of policies may lead to exclusion from service.
We will update this list of policies over time. Larger changes will be announced through the mailing list.
"},{"location":"admin/policies/#cluster-etiquette","title":"Cluster Etiquette","text":"getent passswd $USER to find out the user's office contact details).conda, archive management tools such as tar, (un)zip, or gzip. You should probably only run screen/tmux and maybe a text editor there.hpc-transfer-1 and hpc-transfer-2.In the case of violations marked with a shield () administration reserves the right to remove write and possibly read permission to the given locations. Policies marked with a robot () are automatically enforced.
home, work, and scratch volume). You can request an increase by an email to hpc-helpdesk@bih-charite.de for groups and projects.home 10k files, 1GB spacework 2M files, 1TB spacescratch 20M files, 200TB spacehpc-users and mode is u=rwx,go=; POSIX ACLs are prohibited. This policy is automatically enforced every 5 minutes.u=rwx,g=rwxs,o=; POSIX ACLs are prohibited. This policy is automatically enforced every 5 minutes.scratch/BIH_TRASH after 14 days (by mtime) over night. Trash directories will be removed after 14 further days.touch on files in scratch and subsequently bumping the mtime./tmp). In the case that users need to delete files that they can access but not update/delete, administration will either give write permissions to the Unix group of the work group or project or change the owner to the owner/delegate of this group. This can occur in a group/project directory of a user who has left the organization. In the case that a user leaves the organization, the owner/delegate of the hosting group can request getting access to the user's files with the express agreement of this user./tmp in Slurm-controlled jobs. This will enforce that Slurm can clean up after you.Network connections are a topic important in security. In the case of violations marked with a shield () administration reserves the right to terminate connections without notice and perform other actions.
screen and tmux are only allowed to run on the head nodes. They will be terminated automatically on the compute nodes.srun).~/.ssh/authorized_keys file but their usage is discouraged.~/.bashrc Guide","text":"You can find the current default content of newly created user homes in /etc/skel.bih:
hpc-login-1:~$ head /etc/skel.bih/.bash*\n==> /etc/skel.bih/.bash_logout <==\n# ~/.bash_logout\n\n==> /etc/skel.bih/.bash_profile <==\n# .bash_profile\n\n# Get the aliases and functions\nif [ -f ~/.bashrc ]; then\n . ~/.bashrc\nfi\n\n# User specific environment and startup programs\n\nPATH=$PATH:$HOME/.local/bin:$HOME/bin\n\n==> /etc/skel.bih/.bashrc <==\n# .bashrc\n\n# Source global definitions\nif [ -f /etc/bashrc ]; then\n . /etc/bashrc\nfi\n\n# Uncomment the following line if you don't like systemctl's auto-paging feature:\n# export SYSTEMD_PAGER=\nThis document contains a few tips for helping you using environment modules more effectively. As the general online documentation is lacking a bit, we also give the most popular commands here.
"},{"location":"best-practice/env-modules/#how-does-it-work","title":"How does it Work?","text":"Environment modules are descriptions of software packages. The module command is provided which allows the manipulation of environment variables such as PATH, MANPATH, etc., such that programs are available without passing the full path. Environment modules also allow specifying dependencies between packages and conflicting packages (e.g., when the same binary is available in two packages). Further, environment variables allow the parallel installation of different software versions in parallel and then using software \"a la carte\" in your projects.
List currently loaded modules:
$ module list\nShow all available modules
$ module avail\nLoad one module, make sure to use a specific version to avoid ambiguities.
$ module load Jannovar/0.16-Java-1.7.0_80\nUnload one module
$ module unload Jannovar\nUnload all modules
$ module purge\nGet help for environment modules
$ module help\nGet help for a particular environment module
$ module help Jannovar/0.16-Java-1.7.0_80\nYou can also create your own environment modules. Simply create a directory with module files and then use module use for using the modules from the directory tree.
$ module use path/to/modules\n-bash: module: command not found?","text":"On the login nodes, the module command is not installed. You should not run any computations there, so why would you need environment modules there? ;)
meg-login2$ module\n-bash: module: command not found\nUse srun --pty bash -i to get to one of the compute nodes.
You will certainly finding yourself using a set of programs regularly without it being part of the core cluster installation, e.g., SAMtools, or Python 3. Just putting the appropriate module load lines in your ~/.bashrc will generate warnings when logging into the login node. It is thus recommended to use the following snippet for loading modules automatically on logging into a compute node:
case \"${HOSTNAME}\" in\n login-*)\n ;;\n *)\n # load Python3 environment module\n module load Python/3.4.3-foss-2015a\n\n # Define path for temporary directories, don't forget to cleanup!\n # Also, this will only work after /fast is available.\n export TMPDIR=/data/cephfs-1/home/users/$USER/scratch/tmp\n ;;\nesac\nUnder Construction
This guide was written for the old GPFS file system and is in the process of being updated.
"},{"location":"best-practice/project-structure/#general-aims","title":"General Aims","text":"Mostly, you can separate the files in your projects/pipelines into one of the following categories:
Ideally, scripts and documentation are independent of a given project and can be separated from the rest. Configuration is project-dependent and small and mostly does not contain any sensitive information (such as genotypes that allows for reidentification of donors). In most cases, data might be large and is either also stored elsewhere or together with scripts and configuration can be regenerated easily.
There is no backup of work and scratch
The cluster GPFS file system /fast is not appropriate for keeping around single \"master\" copies of data. You should have a backup and archival strategy for your valuable \"master\" copy data.
In addition, you might need project-specific \"wrapper\" scripts that just call your project-independent script with the correct paths for your project. These scripts rather fall into the \"configuration\" category and should then live together with your configuration.
"},{"location":"best-practice/project-structure/#data","title":"Data","text":"Temporary files
You really should keep temporary files in a temporary directory, set the environment variable TMPDIR appropriately and automatically clean them up (see Useful Tips: Temporary Files)
But how can we put this into practice? Below, we give some examples of how to do this. Note that for simplicity's sake we put all scripts and configuration into one directory/repository contrary to the best practices above. This is for educational purposes only and you should strive for reuseable scripts where it makes sense and separate scripts and configuration.
We will limit this to simple Bash scripts for education's purposes. You should be able to easily adapt this to your use cases.
Thus, the aim is to separate the data from the non-data part of the project such that we can put the non-data part of the project into a separate location and under version control. We call the location for non-data part of the project the home location of your project and the location for the data part of the project the work location of your project.
Overall, we have three options:
Creating the work directory and copy the input files into work/input.
$ mkdir -p project/work/input\n$ cp /data/cephfs-1/work/projects/cubit/tutorial/input/* project/work/input\nCreating the home space. We initialize a Git repository, properly configure the .gitignore file and add a README.md file.
$ mkdir -p project/home\n$ cd project/home\n$ cat <<EOF >.gitignore\n*~\n.*.sw?\nEOF\n$ cat <<EOF >README.md\n# Example Project\n\nThis is an example project with config/scripts linked into work location.\nEOF\n$ git init\n$ git add .gitignore README.md\n$ git commit -m 'Initial project#\nWe then create the a simple script for executing the mapping step and a configuration file that gives the path to the index and list of samples to process.
$ mkdir scripts\n$ cat <<\"EOF\" >scripts/run-mapping.sh\n#!/bin/bash\n\n# Unofficial Bash script mode, see:\n# http://redsymbol.net/articles/unofficial-bash-strict-mode/\nset -euo pipefail\n\n# Get directory to bash file, see\n# https://stackoverflow.com/a/4774063/84349\nSCRIPTPATH=\"$( cd \"$(dirname \"$0\")\" ; pwd -P )\"\n\n# Helper function to print help to stderr.\nhelp()\n{\n >&2 echo \"Run Mapping Step\"\n >&2 echo \"\"\n >&2 echo \"run-mapping.sh [-c config.sh] [-h]\"\n}\n\n# Parse command line arguments into bash variables.\nCONFIG=\nwhile getopts \"hs:\" arg; do\n case $arg in\n h)\n help()\n exit\n ;;\n s)\n CONFIG=$OPTARG\n ;;\n esac\ndone\n\n# Print the executed commands.\nset -x\n\n# Load default configuration, then load configuration file if any was given.\nsource $SCRIPTPATH/../config/default-config.sh\nif [[ -z \"$CONFIG\" ]]; then\n source $CONFIG\nfi\n\n# Create output directory.\nmkdir -p output\n\n# Actually perform the mapping. This assumes that you have\n# made the bwa and samtools commands available, e.g., using conda.\nfor sample in $SAMPLES; do\n bwa mem \\\n $BWA_INDEX \\\n input/${sample}_R1.fq.gz \\\n input/${sample}_R2.fq.gz \\\n | samtools sort \\\n -o output/${sample}.bam \\\n /dev/stdin\ndone\n\nEOF\n$ chmod +x scripts/run-mapping.sh\n$ mkdir -p config\n$ cat <<\"EOF\" >config/default-config.sh\nBWA_INDEX=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/hs37d5/hs37d5.fa\nSAMPLES=\nEOF\n$ cat <<\"EOF\" >config/project-config.sh\n$ BWA_INDEX comes from default configuration already\nSAMPLES=test\nEOF\nThis concludes the basic project setup. Now, to the symlinks:
$ cd ../work\n$ ln -s ../home/scripts ../home/config .\nAnd, to the execution...
$ ./scripts/run-mapping -c config/project-config.sh\n[...]\nWe can reuse the project up to the statement \"This concludes the basic project setup\" in the example for option 1.
Then, we can do the following:
$ cd ../work\n$ mkdir -p output\n\n$ cd ../home\n$ cat <<\"EOF\" >>.gitignore\n\n# Ignore all data\ninput/\nwork/\noutput/\nEOF\n$ git add .gitignore\n$ git commit -m 'Ignoring data file in .gitignore'\n$ ln -s ../work ../output .\nAnd we can execute everything in the home directory.
$ ./scripts/run-mapping -c config/project-config.sh\n[...]\nAgain, we can reuse the project up to the statement \"This concludes the basic project setup\" in the example for option 1.
Then, we do the following:
$ cd ../work\n$ cat <<\"EOF\" >do-run-mapping.sh\n#!/bin/bash\n\n../home/scripts/run-mapping.sh \\\n -c ../home/config/project-config.sh\nEOF\n$ chmod +x do-run-mapping.sh\nNote that the the do-run.sh script could also go into the project-specific Git repository and be linked into the work directory.
Finally, we can run our pipeline:
$ cd ../work\n$ ./do-run-mapping.sh\n[...]\nThe program screen allows you to detach your session from your current login session. So in case you get disconnected your screen session will stay alive.
Hint
You have to reconnect to screen on the machine that you started it. We thus recommend starting it only on the login nodes and not on a compute node.
"},{"location":"best-practice/screen-tmux/#start-and-terminat-a-screen-session","title":"Start and terminat a screen session","text":"You start a new screen session by
$ screen\n$ exit\nIf you want to detach your screen session press Ctrl+a d
To list all your screen sessions run
$ screen -ls\n\nThere is a screen on:\n 2441.pts-1.med0236 (Detached)\n1 Socket in /var/run/screen/S-kbentel.\nTo reattach a screen session run
$ screen -r screen_session_id\nIf you do not know the screen_session_id you can get it with screen -ls, e.g. 2441.pts-1.med0236 in the example above. You do not have to type the whole screen_session_id only as much as is necessary to identify it uniquely. In case there is only one screen session detached it is enough to run screen -r
Sometimes it is necessary to kill a detached screen session. This is done with the command
$ screen -X -S screen_session_id quit\nIt is possible to have multiple windows in a screen session. So suppose you are logged into a screen session, these are the relevant shortcuts
new win: Ctrl+a c\nnext/previous win: Ctrl+a n/p\nTo terminate a window just enter
$ exit\nHere is a sensible screen configuration. Save it as ~/.screenrc.
screenrc
"},{"location":"best-practice/screen-tmux/#fix-a-broken-screen-session","title":"Fix a broken screen session","text":"In case your screen session doesn't write to the terminal correctly, i.e. the formatting of the output is broken, you can fix it by typing to the terminal:
$ tput smam\nComputer software, or simply software, is a generic term that refers to a collection of data or computer instructions that tell the computer how to work, in contrast to the physical hardware from which the system is built, that actually performs the work. -- Wikipedia: Software
As you will most probably never have contact with the HPC system hardware, everything you interact with on the HPC is software. All of your scripts, your configuration files, programs installed by you or administration, and all of your data.
This should also answer the question why you should care about software and why you should try to create and use software of a minimal quality.
Software craftsmanship is an approach to software development that emphasizes the coding skills of the software developers themselves. -- Wikipedia: Software Craftmanship
This Wiki page is not mean to give you an introduction of creating good software but rather collect a (growing) list of easy-to-use and high-impact points to improve software quality. Also, it provides pointers to resources elsewhere on the internet.
"},{"location":"best-practice/software-craftmanship/#use-version-control","title":"Use Version Control","text":"Use a version control system for your configuration and your code. Full stop. Modern version control systems are Git and Subversion.
Every user should have their own Git/Subversion checkout. Otherwise you are inviting a large number of problems.
"},{"location":"best-practice/software-craftmanship/#document-your-code","title":"Document Your Code","text":"This includes
Document where you got things from, how to re-download, etc. E.g., put a README file into each of your data top level directories.
"},{"location":"best-practice/software-craftmanship/#use-checksums","title":"Use Checksums","text":"Use MD5 or other checksums for your data. For example, md5sum and hashdeep are useful utilities for computing and checking them:
md5sum How-To (tools such as sha256sum work the same...)hashdeep How-ToUse some system for managing your workflows. These systems support you by
Snakemake is a popular workflow management system widely used in Bioinformatics. A minimal approach is using Makefiles.
"},{"location":"best-practice/software-craftmanship/#understand-bash-and-shell-exit-codes","title":"Understand Bash and Shell Exit Codes","text":"If you don't want to use a workflow management system, e.g., for one-step jobs, you should at least understand Bash job management and exit codes. For example, you can use if/then/fi in Bash together with exit codes to:
if [[ ! -e file.md5 ]]; then\n md5sum file >file.md5 \\\n || rm -f file.md5\nfi\nAlso, learn about the inofficial Bash strict mode.
"},{"location":"best-practice/software-installation-with-conda/","title":"Software Installation with Conda","text":""},{"location":"best-practice/software-installation-with-conda/#conda","title":"Conda","text":"Users do not have the rights to install system packages on the BIH HPC cluster. For the management of bioinformatics software we therefore recommend using the conda package manager. Conda provides software in different \u201cchannels\u201d and one of those channels contains a huge selection of bioinformatics software (bioconda). Generally packages are pre-compiled and conda just downloads the binaries from the conda servers.
You are in charge of managing your own software stack, but conda makes it easy to do so. We will provide you with a description on how to install conda and how to use it. Of course there are many online resources that you can also use. Please find a list at the end of the document.
Warning
Following a change in their terms of service Anaconda Inc. has started to demand payment from research institutions for using both Anaconda, Miniconda, and the defaults channel. As a consequence, usage of this software is prohibited and we're recommending the alternative free \"miniforge\" distribution instead.
"},{"location":"best-practice/software-installation-with-conda/#premise","title":"Premise","text":"When you logged into the cluster, please make sure that you also executed srun to log into a computation node and perform the software installation there.
hpc-login-1:~$ srun --mem=5G --pty bash -i\nhpc-cpu-123:~$ wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-x86_64.sh\nhpc-cpu-123:~$ bash Miniforge3-Linux-x86_64.sh -b -f -p $HOME/work/miniforge\nhpc-cpu-123:~$ eval \"$(/$HOME/work/miniforge/bin/conda shell.bash hook)\"\nhpc-cpu-123:~$ conda init\nhpc-cpu-123:~$ conda config --set auto_activate_base false\nThis will install conda to $HOME/work/miniforge. You can change the path to your liking, but please note that your $HOME folder has limited space. The work subfolder however has a bigger quota. More about this here.
To make bioinformatics software available, we have to add the bioconda channel to the conda configuration:
hpc-cpu-123:~$ conda config --add channels bioconda\nInstalling packages with conda is straight forward:
hpc-cpu-123:~$ conda install <package>\nThis will install a package into the conda base environment. We will explain environments in detail in the next section. To search for a package, e.g. to find the correct name in conda or if it exists at all, issue the command:
hpc-cpu-123:~$ conda search <string>\nTo choose a specific version (conda will install the latest version that is compatible with the current installed Python version), you can provide the version as follows:
hpc-cpu-123:~$ conda install <package>=<version>\nPlease note that new conda installs may ship with a recently update Python version and not all packages might have been adapted. E.g., if you find out that some packages don't work after starting out/upgrading to Python 3.8, simply try to downgrade Python to 3.7 with conda install python=3.7.
Hint
As resolving the dependency tree of an installation candidate can take a lot of time in Conda, especially when you are installing software from an environment.yaml file, an alternative resolver has been presented that you can use to install software into your Conda environment. The time savings are immense and an installation that took more than an hour can be resolved in seconds.
Simply run
hpc-cpu-123:~$ conda install mamba\nWith that, you can install software into your environment using the same syntax as for Conda:
hpc-cpu-123:~$ mamba install <package>\nConda lets you create environments, such that you can test things in a different environment or group your software. Another common use case is to have different environments for the different Python versions. Since conda is Python-based, conflicting packages will mostly struggle with the Python version.
By default, conda will install packages into its root environment. Please note that software that does not depend on Python and is installed in the root environment, is is available in all other environments.
To create a Python 2.7 environment and activate it, issue the following commands:
hpc-cpu-123:~$ conda create -n py27 python=2.7\nhpc-cpu-123:~$ source activate py27\n(py27) hpc-cpu-123:~$\nFrom now on, conda will install packages into the py27 environment when you issue the install command. To switch back to the root environment, simply deactivate the py27 environment:
(py27) hpc-cpu-123:~$ source deactivate py27\nhpc-cpu-123:~$\nBut of course, as Python 2.7 is not supported any more by the Python Software Foundation, you should switch over to Python 3 already!
"},{"location":"best-practice/temp-files/","title":"Temporary Files","text":"Temporary Files and Slurm
See Slurm: Temporary Files for information how Slurm controls access to local temporary storage.
Often, it is necessary to use temporary files, i.e., write something out in the middle of your program, read it in again later, and then discard these files. For example, samtools sort has to write out chunks of sorted read alignments for allowing to sort files larger than main memory.
TMPDIR","text":"Traditionally, in Unix, the environment variables TMPDIR is used for storing the location of the temporary directory. When undefined, usually /tmp is used.
Generally, there are two locations where you could put temporary files:
/data/cephfs-1/home/users/$USER/scratch/tmp -- inside your scratch folder on the CephFS file system; this location is available from all cluster nodes/tmp -- on the local node's temporary folder; this location is only available on the node itself. The slurm scheduler uses Linux namespaces such that every job gets its private /tmp even when run on the same node.scratch/tmp","text":"Use CephFS-based TMPDIR
Generally setup your environment to use /data/cephfs-1/home/users/$USER/scratch/tmp as filling the local disk of a node with forgotten files can cause a lot of problems.
Ideally, you append the following to your ~/.bashrc to use /data/cephfs-1/home/users/$USER/scratch/tmp as the temporary directory. This will also create the directory if it does not exist. Further, it will create one directory per host name which prevents too many entries in the temporary directory.
export TMPDIR=$HOME/scratch/tmp/$(hostname)\nmkdir -p $TMPDIR\nPrepending this to your job scripts is also recommended as it will ensure that the temporary directory exists.
"},{"location":"best-practice/temp-files/#tmpdir-and-the-scheduler","title":"TMPDIR and the scheduler","text":"In the older nodes, the local disk is a relatively slow spinning disk, in the newer nodes, the local disk is a relatively fast SSD. Further, the local disk is independent from the CephFS file system, so I/O volume to it does not affect the network or any other job on other nodes. Please note that by default, Slurm will not change your environment variables. This includes the environment variable TMPDIR.
Slurm will automatically update temporary files in a job's /tmp on the local file system when the job terminates. To automatically clean up temporary directories on the shared file system, use the following tip.
You can use the following code at the top of your job script to set TMPDIR to the location in your home directory and get the directory automatically cleaned when the job is done (regardless of successful or erroneous completion):
# First, point TMPDIR to the scratch in your home as mktemp will use thi\nexport TMPDIR=$HOME/scratch/tmp\n# Second, create another unique temporary directory within this directory\nexport TMPDIR=$(mktemp -d)\n# Finally, setup the cleanup trap\ntrap \"rm -rf $TMPDIR\" EXIT\nWe recommend to use the program MobaXterm on Windows. MobaXterm is a software that allows you to connect to an SSH server, much like PuTTy, but also maintains your SSH key.
Alternative SSH Clients for Windows
For transfering data from/to Windows, we recommand using WinSCP. Install the latest version from here: https://winscp.net/eng/download.php
On the Login screen of WinSCP create a new login by selecting New Site.
Fill in the following parameters:
File protocol: SFTPHost name: hpc-transfer-1.cubi.bihealth.org or hpc-transfer-2.cubi.bihealth.orgUser name: your user nameGo to Advanced > SSH > Authentication > Authentication parameters > Private key file and select your private ssh key file (in .ppk format).
Press Ok then Save.
Press Login to connect. It will ask for your private key passphrase, if you set one up.
If you need to convert your private ssh key file the .ppk format, on the WinSCP login screen go to Tools > PuTTYgen and follow the steps here: https://docs.acquia.com/cloud-platform/manage/ssh/sftp-key/
Click on Session.
Click on SSH.
In Basic SSH settings, enter a hostname (hpc-login-X.cubi.bihealth.org, where X is 1 or 2), check Specify username and enter your username in the textfield. Select the tab Advanced SSH settings, check Use private key and select your private SSH key file (possible choices described with the next to figures).
Select the id_rsa file generated in Linux OR
select the id_rsa.ppk file generated in Windows with MobaXterm.
Afterwards hit the OK button and MobaXterm will connect.
The session will be stored automatically and you can establish new connections later on, or also multiple ones at the same time, if you like.
"},{"location":"connecting/connecting/","title":"Connecting to HPC 4 Research","text":"HPC 4 Research is only available via the Charit\u00e9, MDC, and BIH internal networks. VPN access requires additional measures which are described in Connecting from External Networks.
There are two primary methods for interacting with BIH HPC:
This part of the documentation only described direct console access via SSH. For information regarding the web portal, please read OnDemand Portal. In case you're not familiar with SSH, you should probably start via the web portal or (if you are determined to learn) read through our SSH basics page.
"},{"location":"connecting/connecting/#in-brief","title":"In brief","text":"Follow these steps to connect to BIH HPC via the command line:
Connect to one of the two login nodes.
# Charite Users\n$ ssh user_c@hpc-login-1.cubi.bihealth.org\n$ ssh user_c@hpc-login-2.cubi.bihealth.org\n\n# MDC Users\n$ ssh user_m@hpc-login-1.cubi.bihealth.org\n$ ssh user_m@hpc-login-2.cubi.bihealth.org\nHint
There are two login nodes, hpc-login-1 and hpc-login-2. There are two for redundancy reasons. Please do not perform big file transfers or an sshfs mount via the login nodes. For this purpose, we have hpc-transfer-1 and hpc-transfer-2.
Please also read Advanced SSH for more custom scenarios how to connect to BIH HPC. If you are using a Windows PC to access BIH HPC, please read Connecting via SSH on Windows
Allocate resources on a computation node using Slurm. Do not compute on the login node!
# Start interactive shell on computation node\n$ srun --pty bash -i\nBonus: Configure your SSH client on Linux and Mac or Windows.
tl;dr
SSH-Based Access:
# Interactive login (choose one)\nssh username@hpc-login-1.cubi.bihealth.org\nssh username@hpc-login-2.cubi.bihealth.org\nsrun --pty bash -i\n\n# File Transfer (choose one)\nsftp local/file username@hpc-transfer-1.cubi.bihealth.org:remote/file\nsftp username@hpc-transfer-2.cubi.bihealth.org:remote/file local/file\n\n# Interactive login into the transfer nodes (choose one)\nssh username@hpc-transfer-1.cubi.bihealth.org\nssh username@hpc-transfer-2.cubi.bihealth.org\nYour username for accessing the cluster are composed of your username at your primary organization (Charit\u00e9/MDC) and a suffix:
<Charite username>_c -> doej_c<MDC username>_m -> jdoe_mPlease read Connecting from External Networks
"},{"location":"connecting/connecting/#i-have-problems-connecting","title":"I have problems connecting","text":"Please read Debugging Connection Problems
"},{"location":"connecting/connection-problems/","title":"Debugging Connection Problems","text":"When you encounter problems with the login to the cluster although we indicated that you should have access, depending on the issue, here is a list of how to solve the problem:
"},{"location":"connecting/connection-problems/#im-getting-a-connection-refused","title":"I'm getting a \"connection refused\"","text":"The full error message looks as follows:
ssh: connect to host hpc-login-1.cubi.bihealth.org port 22: Connection refused\nThis means that your computer could not open a network connection to the server.
<DEST>):ifconfig\ntraceroute <DEST>\nipconfig\ntracepath <DEST>\nYou're logging into BIH HPC cluster! (login-1)\n\n ***Your account has not been granted cluster access yet.***\n\n If you think that you should have access, please contact\n hpc-helpdesk@bih-charite.de for assistance.\n\n For applying for cluster access, contact hpc-helpdesk@bih-charite.de.\n\nuser@login-1's password:\nHint
This is the most common error, and the main cause for this is a wrong username. Please take a couple of minutes to read the What is my username?!
If you encounter this message although we told you that you have access and you checked the username as mentioned above, please write to hpc-helpdesk@bih-charite.de, always indicating the message you get and a detailed description of what you did.
"},{"location":"connecting/connection-problems/#im-getting-a-passphrase-prompt","title":"I'm getting a passPHRASE prompt","text":"You're logging into BIH HPC cluster! (login-1)\n\n *** It looks like your account has access. ***\n\n Login is based on **SSH keys only**, if you are getting a password prompt\n then please contact hpc-helpdesk@bih-charite.de for assistance.\n\nEnter passphrase for key '/home/USER/.ssh/id_rsa':\nHere you have to enter the passphrase that was used for encrypting your private key. Read SSH Basics for further information of what is going on here.
"},{"location":"connecting/connection-problems/#i-can-connect-but-i-get-a-password-prompt","title":"I can connect, but I get a passWORD prompt","text":"You're logging into BIH HPC cluster! (login-1)\n\n *** It looks like your account has access. ***\n\n Login is based on **SSH keys only**, if you are getting a password prompt\n then please contact hpc-helpdesk@bih-charite.de for assistance.\n\nuser@login-1's password:\nThis is diffeerent from passPHRASE prompt
Please see I'm getting a passPHRASE prompt for more information.
When you encounter this message during a login attempt, there is an issue with your SSH key. In this case, please connect with increased verbosity to the cluster (ssh -vvv ...) and mail the output and a detailed description to hpc-helpdesk@bih-charite.de.
If you're using macOS and run into problems that your key is not accepted when connecting, please make sure that you're SSH key is properly unlocked in your SSH agent.
"},{"location":"connecting/from-external/","title":"Connecting from External Networks","text":"This page describes how to connect to the BIH HPC from external networks (e.g., another university or from your home). The options differ depending on your home organization and are described in detail below.
Getting Help with VPN and Gateway Nodes
Please note that the VPNs and gateway nodes are maintained by the central IT departments of Charite/MDC. BIH HPC IT cannot assist you in problems with these serves. Authorative information and documentation is provided by the central IT departments as well.
SSH Key Gotchas
You should use separate SSH key pairs for your workstation, laptop, home computer etc. As a reminder, you will have to register the SSH keys with your home IT organization (MDC or Charite). When using gateway nodes, please make sure to use SSH key agents and agent forwarding (ssh flag \"-A\").
Use the following command to perform a proxy jump via the MDC SSH gateway (ssh1 aka jail1) when connecting to a login node. Note that for logging into the jail, the <MDC_USER> is required.
$ ssh -J <MDC_USER>@ssh1.mdc-berlin.de <HPC_USER>@hpc-login-1.cubi.bihealth.org\nNote
Please Note that the cluster login is independent of access to the MDC jail node ssh1.mdc-berlin.de.
You can find the instructions for getting MDC VPN access here in the MDC intranet below the \"VPN\" heading. Please contact helpdesk@mdc-berlin.de for getting VPN access.
Install the VPN client and then start it. Once VPN has been activated you can SSH to the HPC just as from your workstation.
$ ssh user_m@hpc-login-1.cubi.bihealth.org\nAccess to BIH HPC from external networks (including Eduroam) requires a Charit\u00e9 VPN connection with special access permissions.
"},{"location":"connecting/from-external/#general-charite-vpn-access","title":"General Charit\u00e9 VPN Access","text":"You need to apply for general Charit\u00e9 VPN access if you haven't done so already. The form can be found in the Charite Intranet and contains further instructions. Charit\u00e9 IT Helpdesk can help you with any questions.
"},{"location":"connecting/from-external/#zusatzantrag-b","title":"Zusatzantrag B","text":"Special permissions form B is also required for HPC access. You can find Zusatzantrag B in the Charit\u00e9 intranet. Fill it out and send it to the same address as the general VPN access form above.
Once you have been granted VPN access, start the client and connect to VPN. You will then be able to connect from your client in the VPN just as you do from your workstation.
$ ssh jdoe_c@hpc-login-1.cubi.bihealth.org\nAlternative to using Zusatzantrag B, you can also get access to the Charit\u00e9 VDI (Virtual Desktop Infrastructure). Here, you connect to a virtual desktop computer which is in the Charit\u00e9 network. From there, you can connect to the BIH HPC system.
You need to apply for extended VPN access to be able to access the BIH VDI. The form can be found here. It is important to tick Dienst(e), enter HTTPS and as target view.bihealth.org. Please write to helpdesk@charite.de with the request to access the BIH VDI.
When the access has been set up, follow the instructions on client configuration for Windows, after logging in to the BIH VDI.
"},{"location":"connecting/ssh-basics/","title":"SSH Basics","text":""},{"location":"connecting/ssh-basics/#what-is-ssh","title":"What is SSH?","text":"SSH stands for S ecure Sh ell. It is a software that allows to establish a user-connection to a remote UNIX/Linux machine over the network and remote-control it from your local work-station.
Let's say you have an HPC cluster with hundreds of machines somewhere in a remote data-center and you want to connect to those machines to issue commands and run jobs. Then you would use SSH.
"},{"location":"connecting/ssh-basics/#getting-started","title":"Getting Started","text":""},{"location":"connecting/ssh-basics/#installation","title":"Installation","text":"Simply install your distributions openssh-client package. You should be able to find plenty of good tutorials online. On Windows you can consider using MobaXterm (recommended) or Putty.
Let's call your local machine the client and the remote machine you want to connect to the server.
You will usually have some kind of connection information, like a hostname, IP address and perhaps a port number. Additionally, you should also have received your user-account information stating your user-name, your password, etc.
Follow the instructions below to establish a remote terminal-session.
If your are on Linux
Open a terminal and issue the following command while replacing all the <...> fields with the actual data:
# default port\nssh <username>@<hostname-or-ip-address>\n\n# non-default port\nssh <username>@<hostname-or-ip-address> -p <port-number>\nIf you are on windows
Start putty.exe, go into the Session category and fill out the form, then click the Connect button. Putty also allows to save the connection information in different profiles so you don't have to memorize and retype all fields every time you want to connect.
When you connect to a remote machine via SSH, you will be prompted for your password. This will happen every single time you connect and can feel a bit repetitive at times, especially if you feel that your password is hard to memorize. For those who don't want to type in their password every single time they connect, SSH keys are an alternative way of authentication.
Instead if being prompted for a password, SSH will simply use the key to authenticate. As this key file should be device specific, this also increases security of the login process.
You can generate a new key by issuing:
client:~$ ssh-keygen -t ed25519\n\n# 1. Choose file in which to save the key *(leave blank for default)*\n# 2. Choose a passphrase of at least five characters\nAn SSH key consists of two files, one private and one public key. The public key is installed on remote machines and can only be validated with the matching private key, which is stored on client computers. During the login process this is achieved via public-key cryptography.
Traditionally the algorithm used for this was RSA. Recently elliptic curve cryptography has been developed as a more secure and more performant alternative. We recommend the ed25519 type of SSH key.
The security problem with SSH keys is that anyone with access to the private key has full access to all machines that have the public key installed. Loosing the key or getting it compromised in another way imposes a serious security threat. Therefore, it is best to secure the private key with a passphrase. This passphrase is needed to unlock and use the private key.
Once you have your key-pair generated, you can easily change the passphrase of that key by issuing:
client:~$ ssh-keygen -p\nIn order to avoid having to type the passphrase of the key every time we want to use it, the key can be loaded into an SSH-Agent.
For instance, if you have connected to a login-node via Putty and want to unlock your private key in order to be able to access cluster nodes, you cant configure the SSH-Agent.
client:~$ source <(ssh-agent)\n(The above command will load the required environment variables of the SSH-Agent into your shell environment, effectively making the agent available for your consumption.)
Next, you can load your private key:
client:~$ ssh-add\n(You will be prompted for the passphrase of the key)
You can verify that the agent is running and your key is loaded by issuing:
client:~$ ssh-add -l\n# 'l' as in list-all-loaded-keys\n(The command should print at least one key, showing the key-size, the hash of the key-fingerprint and the location of the file in the file-system.)
Since all home-directories are shared across the entire cluster and you created your key-pair inside your home-directory, you public-key (which is also in your home-directory) is automatically installed on all other cluster nodes, immediately. Try connecting to any cluster node. It should not prompt your for a password.
There is nothing you have to do to \"unload\" or \"lock\" the key-file. Simply disconnect.
"},{"location":"connecting/advanced-ssh/overview/","title":"Advanced SSH usage","text":"Here we describe custom scenarios for using SSH to connect to BIH HPC. To keep it consise, this section is divided into separate documents for
Note
The big Linux distributions automatically manage ssh-agent for you and unlock your keys at login time. If this doesn't work for you, read on.
ssh-agent caches your SSH keys so that you do not need to type your passphrase every time it is used. Activate it by making sure ssh-agent runs in the background and add your key:
$ eval \"$(ssh-agent -s)\"\n$ ssh-add\nor if you chose a custom key name, specify the file like so:
$ ssh-add ~/.ssh/mdc_id_rsa\nIf you run into problems that your key is not accepted when connecting from MacOS, please use:
$ ssh-add --apple-use-keychain\nYou can define a personal SSH configuration file to make connecting to the cluster more comfortable by reducing the typing necessary by a lot. Add the following lines to the file ~/.ssh/config file. Replace USER_NAME with your cluster user name. You can also adapt the Host naming as you like.
Host bihcluster\n HostName hpc-login-1.cubi.bihealth.org\n User USER_NAME\n\nHost bihcluster2\n HostName hpc-login-1.cubi.bihealth.org\n User USER_NAME\nNow, you can do type the following (and you don't have to remember the host name of the login node any more).
$ ssh bihcluster\nThis configuration works if you are inside Charit\u00e9, the Charit\u00e9 VPN, or MDC.
"},{"location":"connecting/advanced-ssh/unix/#mdc-users-jail-node","title":"MDC users: Jail node","text":"If you have an MDC user account and want to connect from the outside, you can use the following ~/.ssh/config lines to set up a ProxyJump via the MDC SSH jail.
Host mdcjail\n HostName ssh1.mdc-berlin.de\n User MDC_USER_NAME\nNow you can run
$ ssh -J mdcjail bihcluster1\nIf you are always connecting from outside the internal network, you can also add a permanent ProxyJump to the SSH configuration like so:
Host bihcluster\n HostName hpc-login-1.cubi.bihealth.org\n User USER_NAME\n ProxyJump mdcjail\nIf you need to connect to the cluster from another computer than the one that contains the SSH keys that you submitted for the cluster login, you have two possibilities.
~/.ssh/id_rsa) to the second computer into the same location.Danger
Do not leave the key on any USB stick. Delete it after file transfer. This is a sensible part of data. Make sure that the files are only readable for you.
$ cd ~/.ssh\n$ chmod g-rwx id_rsa*\n$ ssh-add id_rsa\n$ sshfs <USERNAME>@hpc-transfer-1.cubi.bihealth.org:/ <MOUNTPOINT>\nhpc-transfer-1: follows the structure <host>:<directory> starting in the user home.<MOUNTPOINT> must be an empty but existing and readable directory on your local computerMake sure you have both OSXFUSE and SSHFS installed. You can get both from here: https://osxfuse.github.io/ or the most recent version via Homebrew:
$ brew cask install osxfuse; brew install sshfs; brew link --overwrite sshfs\n$ sshfs -o follow_symlinks <USERNAME>@hpc-transfer-1<X>.cubi.bihealth.org:<directory_relative_to_Cluster_root> <MOUNTPOINT> -o volname=<BIH-FOLDER> -o allow_other,noapplexattr,noappledouble\nDo you really need to run a graphical application on the cluster?
Please note that running more complex Java applications, such as IGV may be not very efficient because of the connection speed. In most cases you can run them on your local workstation by mounting them via SSHFS.
Connect to one of the login nodes using X11 forwarding:
$ ssh -X -C -t <USERNAME>@hpc-login-1.bihealth.org\nOnce you get a login prompt, you can use the srun command with the --x11 parameter to open a X11 session to a cluster node:
$ srun --pty --x11 bash\nAnd finally you can start your X11 application, e.g.:
$ xterm\nAfter a while Visual Terminal should start:
"},{"location":"connecting/advanced-ssh/windows/","title":"Windows","text":""},{"location":"connecting/advanced-ssh/windows/#mounting-the-fs-from-within-the-mdccharite-network","title":"Mounting the FS from within the MDC/Charite Network","text":"Danger
Mounting ssh on Windows is currently discouraged since relevant software is outdated (see also hpc-talk). Also, in most cases it is not really necessary to have a constant mount. For normal data transfer please use WinSCP instead.
Once WinSshFS is started, an icon will be added to your taskbar:
Left-clicking that icon will bring up a window. If not, right click the taskbar icon, select Show Manager and click Add in the menu.
Fill out the marked fields:
hpc-transfer-1.cubi.bihealth.orgPrivateKey. Select the id_rsa private key, not the .ppk format that is provided by PuTTY. Enter the password that you used to secure your key with.Then click Save and then Mount.
Open the explorer. A new drive with the name you gave should show up:
Finished!
"},{"location":"connecting/advanced-ssh/windows/#connecting-via-mdc-jail-node","title":"Connecting via MDC Jail Node","text":"This requires an active MDC account!
Additional to the steps above, click on the tab Network settings.
ssh1.mdc-berlin.de and in the field User your MDC username.Do you really need to run a graphical application on the cluster?
Please note that running more complex Java applications, such as IGV may be not very efficient because of the connection speed. In most cases you can run them on your local workstation by mounting them via SSHFS.
Start MobaXterm, it should automatically fetch your saved Putty sessions as you can see on screen below:
Connect to one of the login nodes, by double-click on saved profile, and then use srun --pty --x11 bash command to start X11 session to one of the nodes:
Finally, start X11 application (below example of starting Visual Terminal):
"},{"location":"connecting/generate-key/unix/","title":"Generating an SSH Key with Unix-like Operating Systems","text":"~/.ssh/id_xxx.pub is present.$ ssh-keygen -t ed25519 -C \"your_email@example.com\"\nWhat is a key passphrase?
You should set a passphrase when generating your key pair. It is used for encrypting your private key in case it is stolen or lost. When using the key for login, you will have to enter the passphrase. Many desktop environments offer ways to automatically unlock your key on login.
Read SSH Basics for more information.
The whole session should look something like this:
host:~$ ssh-keygen -t ed25519 -C \"your_email@example.com\"\nGenerating public/private ed25519 key pair.\nEnter file in which to save the key (/home/USER/.ssh/id_ed25519): \nCreated directory '/home/USER/.ssh'.\nEnter passphrase (empty for no passphrase):\nEnter same passphrase again: \nYour identification has been saved in /home/USER/.ssh/id_ed25519.\nYour public key has been saved in /home/USER/.ssh/id_ed25519.pub.\nThe key fingerprint is:\nSHA256:Z6InW1OYt3loU7z14Kmgy87iIuYNr1gJAN1tG71D7Jc your_email@example.com\nThe key's randomart image is:\n+--[ED25519 256]--+\n|.. . . o |\n|. . . + + |\n|. . = . . |\n|. . +oE. |\n|. So= o o |\n| . . . * = + + |\n| + o + B o o .|\n| oo+. .B + + . |\n|.ooooooo*. . |\n+----[SHA256]-----+\nThe file content of ~/.ssh/id_ed25519.pub should look something like this):
ssh-ed25519 AAAAC3NzaC1lZDI1NTE5AAAAIFzuiaSVD2j5y6RlFxOfREB/Vbd+47ABlxF7du5160ZH your_email@example.com\nAs a next step you need to submit the SSH key use these links as:
Prerequisite: Installing an SSH Client
Please install an SSH client for Windows first.
"},{"location":"connecting/generate-key/windows/#generate-the-key","title":"Generate the Key","text":"Click on Tools and MobaKeyGen (SSH key generator)
In the section Parameters make sure to set the following properties:
RSA (this is the SSH-2 protocol)4096If all is set, hit the Generate button.
During generation, move the mouse cursor around in the blank area.
When finished, make sure to protect your generated key with a passphrase. Save the private and public key. The default name under Linux for the public key is id_rsa.pub and id_rsa for the private key, but you can name them however you want (the .pub is NOT automatically added). Note that in the whole cluster wiki we will use this file naming convention. Also note that the private key will be stored in Putty format (.ppk, this extension is added automatically).
What is your key's passphrase?
You should set a passphrase when generating your private key. This passphrase is used for encrypting you private key to protect it against the private key file theft/being lost. When using the key for login, you will have to enter it (or the first time you load it into the SSH key agent). Note that when being asked for the passphrase this does not occur on the cluster (and is thus unrelated to it) but on your local computer.
Also see SSH Basics for more information.
The gibberish in the textbox is your public key in the format how it has to be submitted to the MDC and Charite (links for this step below). Thus, copy this text and paste it to the SSH-key-submission-web-service of your institution.
Store the private key additionally in the OpenSSH format. To do so, click Conversions and select Export OpenSSH key. To be consistent, give the file the same name as your .ppk private key file above (just without the .ppk).
To summarize, you should end up with three files:
id_rsa.pub The public key file, it is not required if you copy and submit the SSH public key as described above and in the links below.id_rsa.ppk This file is only needed if you plan to use Putty.id_rsa This is your private key and the one and only most important file to access the cluster. It will be added to the sessions in MobaXterm and WinSSHFS (if required).As a next step you need to submit the SSH key use these links as:
As of February 2020, SSH key submission not accepted via email anymore. Instead, use the process outline here.
For any help, please contact helpdesk@charite.de (as this site is maintained by Charite GB IT).
"},{"location":"connecting/submit-key/charite/#charite-zugangsportal","title":"Charite Zugangsportal","text":"Key are submitted in the Charite Zugangsportal. As of Feb 4, you have to use the \"test\" version for this.
Go to zugang.charite.de and login.
Follow through the login page until you reach the main menu (it's tedious but we belive in you ;) Click the \"SSH Keys\" button.
Paste your SSH key (starting with ssh-rsa) and ending with the label (usually your email, e.g., john.doe@charite.de) into the box (1) and press append (2). By default, the key can be found in the file ~/.ssh/id_rsa.pub in Linux. If you generated the key in Windows, please paste the copied key from the text box. Repeat as necessary. Optionally, go back to the main menu (3) when done.
If you have generated your SSH key with PuTTy, you must right click on the ppk-file, then choose \"Edit with PuTTYgen\" in the right click menu. Enter your passphrase. Then copy the SSH key out of the upper box (already highlighted in blue).
Check if the key has been added
After you clicked append, your key will be printed back to you (as shown in the blurred picture above).
If your key is not printed back to you then adding the SSH key to zugang.charite.de was not successful. In this case please contact helpdesk@charite.de for assistance as they (Charite GB IT) maintains that system and it is out of our (BIH HPC IT) control.
Once your key has been added, it will take a few minutes for the changes to go live.
"},{"location":"connecting/submit-key/mdc/","title":"Submitting an SSH Key to MDC","text":"For MDC users, SSH keys are submitted through the MDC PersDB interface (see below). PersDB is not maintained by BIH HPC IT but by MDC IT.
Warning
The SSH keys are only activated over night (but automatically). This is out of our control. Contact helpdesk@mdc-berlin.de for more information.
"},{"location":"connecting/submit-key/mdc/#detour-using-mdc-vmware-view-to-get-into-mdc-intranet","title":"Detour: Using MDC VMWare View to get into MDC Intranet","text":"In case you are not within the MDC network, connect to MDC VMWare view first and use the web brower in the Window session.
~/.ssh/id_rsa.pub into the clipboard window. Ensure that the whole file contents is there (should end with your email address). If you generated the key in Windows, please paste the copied key from the text box.Thus, you will only be able to connect the next day. - Bask in the glory of having completed this process.
"},{"location":"cubit/","title":"Overview","text":"The static data installation can be found at /data/cephfs-1/work/projects/cubit/18.12/static_data.
The static data directory contains a sub-directory for the genomes, the precomputed index files for several different popular mapping tools and associated annotation (GFF and GTF files) from Ensembl and GENCODE for each of the available genomes. The top-level directory structure is as follows:
static_data/annotationsapp_supportdbexome_panelexon_listprecomputedreferenceThe following Ensembl and GENCODE versions corresponding to the indicated reference genomes will be made available on the cluster.
Database Version Reference Genome Ensembl 65 NCBIM37 (Ensembl release corresponding to GENCODE M1) Ensembl 67 NCBIM37 (Ensembl release for sanger mouse genome assembly) Ensembl 68 GRCm38 (Ensembl release for sanger mouse genome assembly) Ensembl 74 GRCh37 (Ensembl release for GENCODE 19) Ensembl 75 GRCh37 (Latest release for GRCh37) Ensembl 79 GRCh38 (Ensembl release for GENCODE 22) Ensembl 80 GRCh38 (Ensembl release corresponding to GENCODE 22) Ensembl 80 GRCm38 (Ensembl release corresponding to GENCODE M1) GENCODE M1 NCBIM37 (No gff3 file) GENCODE M5 GRCm38 GENCODE 19 current for GRCh37 GENCODE 22 current for GRCh38The annotation files associated with the indicated genomes can be accessed in the following directories:
static_data/annotation\n\u251c\u2500\u2500 ENSEMBL\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 65\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 67\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 68\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCm38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 74\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 75\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 79\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 80\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCm38\n\u2514\u2500\u2500 GENCODE\n \u251c\u2500\u2500 19\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n \u251c\u2500\u2500 22\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh38\n \u251c\u2500\u2500 M1\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n \u2514\u2500\u2500 M5\n \u2514\u2500\u2500 GRCm38\nThe static_data/app_support directory contains all data files that are shipped with a software package installed in cubit. For blast this is not complete and more databases can be added upon request.
static_data/app_support\n\u251c\u2500\u2500 blast\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 variable\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 nt\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 refseq_protein\n\u251c\u2500\u2500 Delly\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.6.5\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.6.7\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.7.1\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.7.2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.7.3\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 0.7.5\n\u251c\u2500\u2500 GATK_bundle\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 2.8\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 b37\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 hg19\n\u251c\u2500\u2500 Jannovar\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.14\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 0.16\n\u251c\u2500\u2500 kraken\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 0.10.5-cubi20160426\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 bacvir\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 minikraken_20141208\n\u251c\u2500\u2500 Oncotator\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 v1_ds_Jan262015\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 1000genome_db\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 achilles\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cancer_gene_census\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 ccle_by_gene\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 ccle_by_gp\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 clinvar\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cosmic\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cosmic_fusion\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 cosmic_tissue\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 dbNSFP_ds\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 dbsnp\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 dna_repair_genes\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 esp6500SI_v2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 esp6500SI_v2_coverage\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 familial\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 gencode_out2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 gencode_xrefseq\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hgnc\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mutsig\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 oreganno\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 override_lists\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 ref_hg\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 simple_uniprot\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 so_terms\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 tcgascape\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 tumorscape\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 uniprot_aa_annotation\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 uniprot_aa_xform\n\u2514\u2500\u2500 SnpEff\n \u2514\u2500\u2500 4.1\n \u2514\u2500\u2500 data\n \u251c\u2500\u2500 GRCh37.75\n \u251c\u2500\u2500 GRCh38.79\n \u251c\u2500\u2500 GRCm38.79\n \u251c\u2500\u2500 hg19\n \u251c\u2500\u2500 hg38\n \u2514\u2500\u2500 mm10\nThe file formats in the static_data/db folder are mostly .vcf or .bed files. We provide the following databases:
The directory structure is as follows:
static_data/db\n\u251c\u2500\u2500 COSMIC\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 v72\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n\u251c\u2500\u2500 dbNSFP\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 2.9\n\u251c\u2500\u2500 dbSNP\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 b128\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 mm9\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 b142\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 b144\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 b147\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh38\n\u251c\u2500\u2500 DGV\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 2015-07-23\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n\u251c\u2500\u2500 ExAC\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 release0.3\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 release0.3.1\n\u251c\u2500\u2500 giab\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 NA12878_HG001\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 NISTv2.19\n\u251c\u2500\u2500 goldenpath\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 variable\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mm9\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n\u251c\u2500\u2500 SangerMouseGenomesProject\n\u2502 \u2514\u2500\u2500 REL-1211-SNPs_Indels\n\u2502 \u251c\u2500\u2500 mm9\n\u2502 \u2514\u2500\u2500 NCBIM37\n\u2514\u2500\u2500 UK10K_cohort\n \u2514\u2500\u2500 REL-2012-06-02\nThese exome panel data are proprietary and downloaded after registration. In case you want to use them, be sure you have access to them by creating an account at Agilent or Roche to not run into legal trouble.
static_data/exome_panel\n\u251c\u2500\u2500 Agilent\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 SureSelect_Human_All_Exon_V4\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 SureSelect_Human_All_Exon_V5\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 SureSelect_Human_All_Exon_V6\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 SureSelect_Mouse_All_Exon_V1\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mm9\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 NCBIM37\n\u2514\u2500\u2500 Roche\n \u2514\u2500\u2500 SeqCap_EZ_MedExome\n \u2514\u2500\u2500 GRCh37\nHere we provide exon lists for some human genome assemblies in the .bed-file format. Each file exists with the original coordinates contained and as a version with 10 bp padded on each site (suffix: _plus_10bp.bed). The folder structure is self-explanatory:
static_data/exon_list\n\u251c\u2500\u2500 CCDS\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 15\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u2502\u00a0\u00a0 \u2514\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 18\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 hg38\n\u2514\u2500\u2500 ENSEMBL\n \u251c\u2500\u2500 74\n \u2502\u00a0\u00a0 \u2514\u2500\u2500 GRCh37\n \u2514\u2500\u2500 75\n \u2514\u2500\u2500 GRCh37\nIndex files for
have been precomputed. The index corresponding to each genome is stored in the following directory structure with the above mentioned reference genomes as subfolders (listed here only for Bowtie/1.1.2, same subfolders for the remaining programs):
static_data/precomputed\n\u251c\u2500\u2500 Bowtie\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 1.1.2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 danRer10\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 dm6\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 ecoli\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 GRCm38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hg18\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hg38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mm10\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 mm9\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 NCBIM37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 phix\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 sacCer3\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 UniVec\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 UniVec_Core\n\u251c\u2500\u2500 Bowtie2\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 2.2.5\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 [see Bowtie/1.1.2]\n\u251c\u2500\u2500 BWA\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 0.7.12\n\u2502\u00a0\u00a0 \u2502 \u2514\u2500\u2500 [see Bowtie/1.1.2]\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 0.7.15\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 [see Bowtie/1.1.2]\n\u2514\u2500\u2500 STAR\n \u2514\u2500\u2500 2.4.1d\n \u2514\u2500\u2500 default\n \u00a0\u00a0 \u2514\u2500\u2500 [see Bowtie/1.1.2]\nWe provide the NCBI mouse reference assembly used by the Sanger Mouse Genomics group for NCBIM37 and GRCm38. This is a reliable source where the appropriate contigs have already been selected by experts. NCBIM37 is annotated with Ensembl release 67 and GRCm38 with Ensembl release 68.
"},{"location":"cubit/references/#ucsc-mouse-reference-genome-assemblies","title":"UCSC mouse reference genome assemblies","text":"The assembly sequence is in one file per chromosome and is available for mm9 and mm10. We concatenated all the chromosome files to one final fasta file for each genome assembly.
"},{"location":"cubit/references/#ncbi-human-reference-genome-assemblies","title":"NCBI human reference genome assemblies","text":"The assembly sequence is in one file per chromosome is available for hg18, hg19 and hg38. We concatenated all the chromosome files to one final fasta file for each genome assembly. Additionally, in the subfolder chromosomes we keep the chromosome fasta files separately for hg18 and hg19.
The following directory structure indicates the available genomes. Where there isn't a name for the data set, either the source (e.g. sanger - from the Sanger Mouse Genomes project) or the download date is used to name the sub-directory.
static_data/reference\n\u251c\u2500\u2500 danRer10\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 dm6\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 ecoli\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 GCA_000005845.2_ASM584v2\n\u251c\u2500\u2500 genomemedley\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 1\n\u251c\u2500\u2500 GRCh37\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 g1k_phase1\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 g1k_phase2\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hs37\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 hs37d5\n\u251c\u2500\u2500 GRCh38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hs38\n\u2502\u00a0\u00a0 \u251c\u2500\u2500 hs38a\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 hs38DH\n\u251c\u2500\u2500 GRCm38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 sanger\n\u251c\u2500\u2500 hg18\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 hg19\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 hg38\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 mm10\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 mm9\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 NCBIM37\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 sanger\n\u251c\u2500\u2500 phix\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 illumina\n\u251c\u2500\u2500 sacCer3\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 ucsc\n\u251c\u2500\u2500 UniVec\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 9\n\u2514\u2500\u2500 UniVec_Core\n \u2514\u2500\u2500 9\nPlease see the section Connection Problems.
"},{"location":"help/faq/#connecting-to-the-cluster-takes-a-long-time","title":"Connecting to the cluster takes a long time.","text":"The most probable cause for this is a conda installation which loads files on login. To disable this behaviour we can put the conda intialisation code behind a bash alias to run it manually later:
In your ~/.bashrc find the conda block:
# >>> conda initialize >>>\n# !! Contents within this block are managed by 'conda init' !!\n...\n# <<< conda initialize <<<\nEncapsulate the entire section in a bash function like this:
conda_init() {\n # >>> conda initialize >>>\n # !! Contents within this block are managed by 'conda init' !!\n ...\n # <<< conda initialize <<<\n}\nFrom now on to use conda you must first run conda_init which then loads the necessary files.
You can also run the bash shell in verbose mode to find out exactly which command is slowing down login:
$ ssh user@hpc-login-1.cubi.bihealth.org bash -iv\nAdministrativa
Request for both systems are handled separately, depending on the user's affiliation with research/service groups.
Hardware and Systems
Bioinformatics Software
packet_write_wait: Connection to XXX : Broken pipe\". How can I fix this?","text":"Try to put the following line at the top of your ~/.ssh/config.
ServerAliveInterval 30\nThis will make ssh send an empty network package to the server. This will prevent network hardware from thinking your connection is unused/broken and terminating it.
If the problem persists, please report it to hpc-helpdesk@bih-charite.de.
"},{"location":"help/faq/#my-job-terminated-before-being-done-what-happened","title":"My job terminated before being done. What happened?","text":"First of all, look into your job logs. In the case that the job was terminated by Slurm (e.g., because it ran too long), you will find a message like this at the bottom. Please look at the end of the last line in your log file.
slurmstepd: error: *** JOB <your job id> ON med0xxx CANCELLED AT 2020-09-02T21:01:12 DUE TO TIME LIMIT ***\nThis indicates that you need to need to adjust the --time limit to your sbatch command.
slurmstepd: error: Detected 2 oom-kill event(s) in step <your job id>.batch cgroup.\nSome of your processes may have been killed by the cgroup out-of-memory handler\nThis indicates that your job tries to use more memory than has been allocated to it. Also see Slurm Scheduler: Memory Allocation
Otherwise, you can use sacct -j JOBID to read the information that the job accounting system has recorded for your job. A job that was canceled (indicated by CANCELED) by the Slurm job scheduler looks like this (ignore the COMPLETED step that is just some post-job step added by Slurm automatically).
# sacct -j _JOBID_\n JobID JobName Partition Account AllocCPUS State ExitCode\n------------ ---------- ---------- ---------- ---------- ---------- --------\n_JOBID_ snakejob.+ medium hpc-ag-xx+ 4 TIMEOUT 0:0\n_JOBID_.bat+ batch hpc-ag-xx+ 4 CANCELLED 0:15\n_JOBID_.ext+ extern hpc-ag-xx+ 4 COMPLETED 0:0\nUse the --long flag to see all fields (and probably pipe it into less as: sacct -j JOBID --long | less -S). Things to look out for:
MaxRSS)?Elapsed)?Note that --long does not show all fields. For example, the following tells us that the given job was above its elapsed time which caused it to be killed.
# sacct -j _JOBID_ --format Timelimit,Elapsed\n Timelimit Elapsed\n---------- ----------\n 01:00:00 01:00:12\n 01:00:13\n 01:00:12\nUse man sacct, sacct --helpformat, or see the Slurm Documentation for options for the --format field of sacct.
This is most probably caused by your job being allocated insufficient memory. Please see the memory part of the answer to My job terminated before being done. What happened?
"},{"location":"help/faq/#how-can-i-create-a-new-project","title":"How can I create a new project?","text":"You can create a project if you are either a group leader of an AG or a delegate of an AG. If this is the case, please follow these instructions.
"},{"location":"help/faq/#i-cannot-create-pngs-in-r","title":"I cannot create PNGs in R","text":"For using the png method, you need to have an X11 session running. This might be the case if you logged into a cluster node using srun --x11 if configured correctly but is not the case if you submitted a bash job. The solution is to use xvfb-run (xvfb = X11 virtual frame-buffer).
Here is the content of an example script:
$ cat img.R\n#!/usr/bin/env Rscript\n\npng('cars.png')\ncars <- c(1, 3, 6, 4, 9)\nplot(cars)\ndev.off()\nHere, it fails without X11:
$ ./img.R\nError in .External2(C_X11, paste(\"png::\", filename, sep = \"\"), g$width, :\n unable to start device PNG\nCalls: png\nIn addition: Warning message:\nIn png(\"cars.png\") : unable to open connection to X11 display ''\nExecution halted\nHere, it works with xvfb-run:
$ xvfb-run ./img.R\nnull device\n 1\n$ ls\ncars.png foo.png img.R Rplots.pdf\nYou can use scontrol show job JOBID to get the details displayed about your jobs. In the example below, we can see that the job is in the PENDING state. The Reason field tells us that the job did not scheduled because the specified dependency was neverfulfilled. You can find a list of all job reason codes in the Slurm squeue documentation.
JobId=863089 JobName=pipeline_job.sh\n UserId=holtgrem_c(100131) GroupId=hpc-ag-cubi(5272) MCS_label=N/A\n Priority=1 Nice=0 Account=(null) QOS=normal\n JobState=PENDING Reason=DependencyNeverSatisfied Dependency=afterok:863087(failed)\n Requeue=1 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0\n RunTime=00:00:00 TimeLimit=08:00:00 TimeMin=N/A\n SubmitTime=2020-05-03T18:57:34 EligibleTime=Unknown\n AccrueTime=Unknown\n StartTime=Unknown EndTime=Unknown Deadline=N/A\n SuspendTime=None SecsPreSuspend=0 LastSchedEval=2020-05-03T18:57:34\n Partition=debug AllocNode:Sid=hpc-login-1:28797\n ReqNodeList=(null) ExcNodeList=(null)\n NodeList=(null)\n NumNodes=1 NumCPUs=1 NumTasks=1 CPUs/Task=1 ReqB:S:C:T=0:0:*:*\n TRES=cpu=1,node=1,billing=1\n Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*\n MinCPUsNode=1 MinMemoryNode=0 MinTmpDiskNode=0\n Features=(null) DelayBoot=00:00:00\n OverSubscribe=OK Contiguous=0 Licenses=(null) Network=(null)\n Command=/data/cephfs-1/work/projects/medgen_genomes/2019-06-05_genomes_reboot/GRCh37/wgs_cnv_export/pipeline_job.sh\n WorkDir=/data/cephfs-1/work/projects/medgen_genomes/2019-06-05_genomes_reboot/GRCh37/wgs_cnv_export\n StdErr=/data/cephfs-1/work/projects/medgen_genomes/2019-06-05_genomes_reboot/GRCh37/wgs_cnv_export/slurm-863089.out\n StdIn=/dev/null\n StdOut=/data/cephfs-1/work/projects/medgen_genomes/2019-06-05_genomes_reboot/GRCh37/wgs_cnv_export/slurm-863089.out\n Power=\n MailUser=(null) MailType=NONE\nIf you see a Reason=ReqNodeNotAvail,_Reserved_for_maintenance then also see Reservations / Maintenances.
For GPU jobs also see \"My GPU jobs don't get scheduled\".
"},{"location":"help/faq/#my-gpu-jobs-dont-get-scheduled","title":"My GPU jobs don't get scheduled","text":"There are only four GPU machines in the cluster (with four GPUs each, hpc-gpu-1 to hpc-gpu-4). Please inspect first the number of running jobs with GPU resource requests:
hpc-login-1:~$ squeue -o \"%.10i %20j %.2t %.5D %.4C %.10m %.16R %.13b\" \"$@\" | grep hpc-gpu- | sort -k7,7\n 1902163 ONT-basecalling R 1 2 8G hpc-gpu-1 gpu:tesla:2\n 1902167 ONT-basecalling R 1 2 8G hpc-gpu-1 gpu:tesla:2\n 1902164 ONT-basecalling R 1 2 8G hpc-gpu-2 gpu:tesla:2\n 1902166 ONT-basecalling R 1 2 8G hpc-gpu-2 gpu:tesla:2\n 1902162 ONT-basecalling R 1 2 8G hpc-gpu-3 gpu:tesla:2\n 1902165 ONT-basecalling R 1 2 8G hpc-gpu-3 gpu:tesla:2\n 1785264 bash R 1 1 1G hpc-gpu-4 gpu:tesla:2\nThis indicates that there are two free GPUs on hpc-gpu-4.
Second, inspect the node states:
hpc-login-1:~$ sinfo -n hpc-gpu-[1-4]\nPARTITION AVAIL TIMELIMIT NODES STATE NODELIST\ndebug* up 8:00:00 0 n/a\nmedium up 7-00:00:00 0 n/a\nlong up 28-00:00:0 0 n/a\ncritical up 7-00:00:00 0 n/a\nhighmem up 14-00:00:0 0 n/a\ngpu up 14-00:00:0 1 drng hpc-gpu-4\ngpu up 14-00:00:0 3 mix med[0301-0303]\nmpi up 14-00:00:0 0 n/a\nThis tells you that hpc-gpu-1 to hpc-gpu-3 have jobs running (\"mix\" indicates that there are free resources, but these are only CPU cores not GPUs). hpc-gpu-4 is shown to be in \"draining state\". Let's look what's going on there.
hpc-login-1:~$ scontrol show node hpc-gpu-4\nNodeName=hpc-gpu-4 Arch=x86_64 CoresPerSocket=16\n CPUAlloc=2 CPUTot=64 CPULoad=1.44\n AvailableFeatures=skylake\n ActiveFeatures=skylake\n Gres=gpu:tesla:4(S:0-1)\n NodeAddr=hpc-gpu-4 NodeHostName=hpc-gpu-4 Version=20.02.0\n OS=Linux 3.10.0-1127.13.1.el7.x86_64 #1 SMP Tue Jun 23 15:46:38 UTC 2020\n RealMemory=385215 AllocMem=1024 FreeMem=347881 Sockets=2 Boards=1\n State=MIXED+DRAIN ThreadsPerCore=2 TmpDisk=0 Weight=1 Owner=N/A MCS_label=N/A\n Partitions=gpu\n BootTime=2020-06-30T20:33:36 SlurmdStartTime=2020-07-01T09:31:51\n CfgTRES=cpu=64,mem=385215M,billing=64\n AllocTRES=cpu=2,mem=1G\n CapWatts=n/a\n CurrentWatts=0 AveWatts=0\n ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s\n Reason=deep power-off required for PSU [root@2020-07-17T13:21:02]\nThe \"State\" attribute indicates the node has jobs running but is currenlty being \"drained\" (accepts no new jobs). The \"Reason\" gives that it has been scheduled for power-off for maintenance of the power supply unit.
"},{"location":"help/faq/#when-will-my-job-be-scheduled","title":"When will my job be scheduled?","text":"You can use the scontrol show job JOBID command to inspect the scheduling information for your job. For example, the following job is scheduled to start at 2022-09-19T07:53:29 (StartTime) and will be terminated if it does not stop before 2022-09-19T15:53:29 (EndTime) For further information, it has been submitted at 2022-09-15T12:24:57 (SubmitTime) and has been last considered by the scheduler at 2022-09-19T07:53:15 (LastSchedEval).
# scontrol show job 4225062\nJobId=4225062 JobName=C2371_2\n UserId=user_c(133196) GroupId=hpc-ag-group(1030014) MCS_label=N/A\n Priority=805 Nice=0 Account=hpc-ag-group QOS=normal\n JobState=PENDING Reason=QOSMaxCpuPerUserLimit Dependency=(null)\n Requeue=1 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0\n RunTime=00:00:00 TimeLimit=08:00:00 TimeMin=N/A\n SubmitTime=2022-09-15T12:24:57 EligibleTime=2022-09-15T12:24:57\n AccrueTime=2022-09-15T12:24:57\n StartTime=2022-09-19T07:53:29 EndTime=2022-09-19T15:53:29 Deadline=N/A\n SuspendTime=None SecsPreSuspend=0 LastSchedEval=2022-09-19T07:53:15 Scheduler=Main\n Partition=medium AllocNode:Sid=hpc-login-1:557796\n ReqNodeList=(null) ExcNodeList=(null)\n NodeList=(null)\n NumNodes=1-1 NumCPUs=25 NumTasks=25 CPUs/Task=1 ReqB:S:C:T=0:0:*:*\n TRES=cpu=25,mem=150G,node=1,billing=25\n Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*\n MinCPUsNode=1 MinMemoryNode=150G MinTmpDiskNode=0\n Features=(null) DelayBoot=00:00:00\n OverSubscribe=YES Contiguous=0 Licenses=(null) Network=(null)\n Command=/data/cephfs-1/home/users/user_c/work/SCZ_replic/JR_sims/GS_wrapy/wrap_y0_VP_2371_GS_chunk2_C02.sh\n WorkDir=/data/cephfs-1/home/users/user_c/work/SCZ_replic/JR_sims\n StdErr=/data/cephfs-1/home/users/user_c/work/SCZ_replic/JR_sims/E2371_2.txt\n StdIn=/dev/null\n StdOut=/data/cephfs-1/home/users/user_c/work/SCZ_replic/JR_sims/slurm-4225062.out\n Power=\nYou can see the partition that your job runs in with squeue -j JOBID:
hpc-login-1:~$ squeue -j 877092\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 877092 medium snakejob holtgrem R 0:05 1 med0626\nSee Job Scheduler for information about the partition's properties and how jbos are routed to partitions. You can force jobs to run in a particular partition by specifying the --partition parameter, e.g., by adding --partition=medium or -p medium to your srun and sbatch calls.
This is probably answered by the answer to My jobs don't run in the partition I expect.
"},{"location":"help/faq/#how-can-i-mount-a-network-volume-from-elsewhere-on-the-cluster","title":"How can I mount a network volume from elsewhere on the cluster?","text":"You cannot.
"},{"location":"help/faq/#how-can-i-make-workstationserver-files-available-to-the-hpc","title":"How can I make workstation/server files available to the HPC?","text":"You can transfer files to the cluster through Rsync over SSH or through SFTP to the hpc-transfer-1 or hpc-transfer-2 node.
Do not transfer files through the login nodes. Large file transfers through the login nodes can cause performance degradation for the users with interactive SSH connections.
"},{"location":"help/faq/#how-can-i-circumvent-invalid-instruction-signal-4-errors","title":"How can I circumvent \"invalid instruction\" (signal 4) errors?","text":"Make sure that software is compiled with \"sandy bridge\" optimizations and no later one. E.g., use the -march=sandybridge argument to the GCC/LLVM compiler executables.
If you absolutely need it, there are some boxes with more recent processors in the cluster (e.g., Haswell architecture). Look at the /proc/cpuinfo files for details.
Please check whether there might be other jobs waiting in front of you! The following squeue call will show the allocated GPUs of jobs in the gpu queue. This is done by specifying a format string and using the %b field.
squeue -o \"%.10i %9P %20j %10u %.2t %.10M %.6D %10R %b\" -p gpu\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(R TRES_PER_NODE\n 872571 gpu bash user1 R 15:53:25 1 hpc-gpu-3 gpu:tesla:1\n 862261 gpu bash user2 R 2-16:26:59 1 hpc-gpu-4 gpu:tesla:4\n 860771 gpu kidney.job user3 R 2-16:27:12 1 hpc-gpu-2 gpu:tesla:1\n 860772 gpu kidney.job user3 R 2-16:27:12 1 hpc-gpu-2 gpu:tesla:1\n 860773 gpu kidney.job user3 R 2-16:27:12 1 hpc-gpu-2 gpu:tesla:1\n 860770 gpu kidney.job user3 R 4-03:23:08 1 hpc-gpu-1 gpu:tesla:1\n 860766 gpu kidney.job user3 R 4-03:23:11 1 hpc-gpu-3 gpu:tesla:1\n 860767 gpu kidney.job user3 R 4-03:23:11 1 hpc-gpu-1 gpu:tesla:1\n 860768 gpu kidney.job user3 R 4-03:23:11 1 hpc-gpu-1 gpu:tesla:1\nIn the example above, user1 has one job with one GPU running on hpc-gpu-3, user2 has one job running with 4 GPUs on hpc-gpu-4 and user3 has 7 jobs in total running of different machines with one GPU each.
"},{"location":"help/faq/#how-can-i-access-graphical-user-interfaces-such-as-for-matlab-on-the-cluster","title":"How can I access graphical user interfaces (such as for Matlab) on the cluster?","text":"-X for Linux/Mac sshsrun --pty --x11 bash -i (instead of srun --pty --x11 bash -i).Also see:
This is sometimes useful, e.g., for monitoring the CPU/GPU usage of your job interactively.
No Computation Outside of Slurm
Do not perform any computation outside of the scheduler as (1) this breaks the purpose of the scheduling system and (2) administration is not aware and might kill you jobs.
The answer is simple, just SSH into this node.
hpc-login-1:~$ ssh hpc-cpu-xxx\nClassically, jobs on HPC systems are written in a way that they can run on multiple nodes at once, using the network to communicate. Slurm comes from this world and when allocating more than one CPU/core, it might allocate them on different nodes. Please use --nodes=1 to force Slurm to allocate them on a single node.
You can select the CPU architecture by using the -C/--constraint flag to sbatch and srun. The following are available (as detected by the Linux kernel):
ivybridge (96 nodes, plus 4 high-memory nodes)haswell (16 nodes)broadwell (112 nodes)skylake (16 nodes, plus 4 GPU nodes)You can specify contraints with OR such as --constraint=haswell|broadwell|skylake. You can see the assignment of architectures to nodes using the sinfo -o \"%8P %.5a %.10l %.6D %.6t %10f %N\" command. This will also display node partition, availability etc.
No worries!
As documented in the Storage Locations section, each user/project/group has three storage volumes: A small home, a larger work and a large (but temporary) scratch. There are limits on the size of these volumes. You get a nightly warning email in case you are over the soft limit and you will not be able to write any more data if you get above the hard limit. When you login to the login nodes, the quotas and current usage is displayed to you.
Please note that not all files will be displayed when using ls. You have to add the -a parameter to also show files and directory starting with a dot. Often, users are confused if these dot directories take up all of their home quota.
Use the following command to list all files and directories in your home:
hpc-login-1:~$ ls -la ~/\nFor more information on how to keep your home directory clean and avoid quota warnings, please read Home Folder Quota.
"},{"location":"help/faq/#im-getting-a-disk-quota-exceeded-error","title":"I'm getting a \"Disk quota exceeded\" error.","text":"Most probably you are running into the same problem as described above: Help, I'm getting a Quota Warning Email!
"},{"location":"help/faq/#environment-modules-dont-work-and-i-get-module-command-not-found","title":"Environment modules don't work and I get \"module: command not found\"","text":"First of all, ensure that you are on a compute node and not on one of the login nodes. One common reason is that the system-wide Bash configuration has not been loaded, try to execute source /etc/bashrc and then re-try using module. In the case that the problem persists, please contact hpc-helpdesk@bih-charite.de.
All users get their home directory setup using a skelleton files. These file names start with a dot . and are hidden when you type ls, you have to type ls -a to see them. You can find the current skelleton in /etc/skel.bih and inspect the content of the Bash related files as follows:
hpc-login-1:~$ head /etc/skel.bih/.bash*\n==> /etc/skel.bih/.bash_logout <==\n# ~/.bash_logout\n\n==> /etc/skel.bih/.bash_profile <==\n# .bash_profile\n\n# Get the aliases and functions\nif [ -f ~/.bashrc ]; then\n . ~/.bashrc\nfi\n\n# User specific environment and startup programs\n\nPATH=$PATH:$HOME/.local/bin:$HOME/bin\n\n==> /etc/skel.bih/.bashrc <==\n# .bashrc\n\n# Source global definitions\nif [ -f /etc/bashrc ]; then\n . /etc/bashrc\nfi\n\n# Uncomment the following line if you don't like systemctl's auto-paging feature:\n# export SYSTEMD_PAGER=\nThere actually are a couple of more files by default. The original copy in /etc/skel.bih might slightly change over time during improvements but we will not touch your home directory in an unsolicited way at any time!
hpc-login-1:~$ tree -a /etc/skel.bih/\n/etc/skel.bih/\n\u251c\u2500\u2500 .bash_logout\n\u251c\u2500\u2500 .bash_profile\n\u251c\u2500\u2500 .bashrc\n\u251c\u2500\u2500 .screenrc\n\u2514\u2500\u2500 .vimrc\nHave a look at our How-To: Debug Software and How-To: Debug Software on HPC Systems guides!
But it works on my workstation!
Yes, please also refer to these guides on possible approaches to find the problem.
"},{"location":"help/faq/#my-r-sessions-is-always-crashing","title":"My R sessions is always crashing!","text":"We are aware of issues for macOS users that are related to locales. You might see something like the following in your R session or logs.
Note that this is not macOS specific by itself but by macOS using a locale that R does not understand.
Try using export LC_ALL=C before calling R. If the problem goes away, you can add the following lines to your ~/.bashrc file to make this permanent. Note that this is actually a good idea so your locale settings are fixed.
# Fix all locale settings to \"C\" for robustness.\nexport LC_ALL=C\nChatGPT on Unix Locales
Let's se what ChatGPT 3.5 tells us about Unix locale.
Me: Please explain the concept of Unix locales to me.
Certainly! Unix locales are a set of environment variables and settings that define how various aspects of software and user interactions should be handled to accommodate different languages, regions, and cultural conventions. They play a crucial role in enabling software to adapt to local preferences for displaying text, formatting numbers, handling dates, and more.
In a Unix-like operating system (such as Linux or macOS), locales provide a way to customize how programs behave based on the user's language, country, and cultural norms. Here are some of the key components of Unix locales:
Language: The primary language used for messages, menus, prompts, and system output. For instance, if you set your locale to \"en_US\", messages and interfaces will be in English (United States).
Region: The country or region that defines conventions for aspects like date and time formatting, number formatting (decimal point or comma, thousands separator, etc.), and currency symbols. For example, \"en_GB\" represents English (Great Britain).
Character Encoding: Specifies how characters are encoded, which is vital for correctly displaying and handling text in various scripts and languages. Common encodings include UTF-8 for Unicode text.
Collation: Defines how strings should be sorted. Different languages have different sorting orders for characters, and the locale determines the appropriate collation rules.
Numeric and Date Formats: Specifies how numbers, dates, and times should be formatted. Different regions have different conventions for date formats (dd/mm/yyyy vs. mm/dd/yyyy) and numerical representations (commas vs. periods as decimal points).
Currency Symbols: Defines the symbol or abbreviation used for representing currency. This is important for currency formatting in software.
Unix locales are typically identified by a combination of language and region codes, like \"en_US\" for American English or \"fr_FR\" for French (France). These locales are stored in files on the system, and they contain the necessary rules and data for proper localization.
Users and administrators can set the system's default locale, and individual users can override this setting in their environment. Here are some common commands used to manage locales:
locale: Displays the current locale settings.locale -a: Lists all available locales on the system.export LC_ALL=locale_code: Temporarily sets the locale for the current session.update-locale locale_code: Changes the system-wide default locale.Programs that are designed to be locale-aware use these settings to display messages, format numbers and dates, and handle other language-dependent features. This allows software to be more user-friendly and relevant to different linguistic and cultural contexts.
In summary, Unix locales provide a framework for adapting software behavior to match the language, region, and cultural preferences of users, ensuring a more inclusive and globally accessible computing experience.
"},{"location":"help/faq/#which-cuda-version-is-installed","title":"Which CUDA version is installed?","text":"For this, connect to the node you want to query (via SSH but do not perform any computation via SSH!)
hpc-login-1:~$ ssh hpc-gpu-1\nhpc-gpu-1:~$ yum list installed 2>/dev/null | grep cuda.x86_64\ncuda.x86_64 10.2.89-1 @local-cuda\nnvidia-driver-latest-dkms-cuda.x86_64 3:440.64.00-1.el7 @local-cuda\nNo, as Docker essentially gives you access as the root user.
However, you can use Apptainer (former Singularity) to run containers (and even many Docker contains if they are \"properly built\"). Also see Using Apptainer (with Docker Images).
"},{"location":"help/faq/#how-can-i-copy-data-between-the-max-cluster-mdc-network-and-bih-hpc","title":"How can I copy data between the MAX Cluster (MDC Network) and BIH HPC?","text":"The MAX cluster is the HPC system of the MDC. It is located in the MDC network. The BIH HPC is located in the BIH network.
In general, connections can only be initiated from the MDC network to the BIH network. The reverse does not work. In other words, you have to log into the MAX cluster and then initiate your file copies to or from the BIH HPC from there. E.g., use rsync -avP some/path user_m@hpc-transfer-1.cubi.bihealth.org:/another/path to copy files from the MAX cluster to BIH HPC and rsync -avP user_m@hpc-transfer-1.cubi.bihealth.org:/another/path some/path to copy data from the BIH HPC to the MAX cluster.
In general, connections can only be initiated from the Charit\u00e9 network to the BIH network. The reverse does not work. In other words, you have to be on a machine inside the Charit\u00e9 network and then initiate your file copies to or from the BIH HPC from there. E.g., use rsync -avP some/path user_c@hpc-transfer-1.cubi.bihealth.org:/another/path to copy files from the MAX cluster to BIH HPC and rsync -avP user_c@hpc-transfer-1.cubi.bihealth.org:/another/path some/path to copy data from the BIH HPC to the MAX cluster.
As of December 3, 2020 we have established a policy to limit you to 512 files and 128MB of RAM. Further, you are limited to using the equivalent of one core. This limit is enforced for all processes originating from an SSH session and the limit is enforced on all jobs. This was done to prevent users from thrashing the head nodes or using SSH based sessions for computation.
"},{"location":"help/faq/#slurm-complains-about-execve-no-such-file-or-directory","title":"Slurm complains aboutexecve / \"No such file or directory\"","text":"This means that the program that you want to execute does not exist. Consider the following example:
[user@hpc-login-1 ~]$ srun --time 2-0 --nodes=1 --ntasks-per-node=1 \\\n --cpus-per-task=12 --mem 96G --partition staging --immediate 5 \\\n --pty bash -i\nslurmstepd: error: execve(): 5: No such file or directory\nsrun: error: hpc-cpu-2: task 0: Exited with exit code 2\nCan you spot the problem? In this case, the problem is that for long arguments such as --mem you must use the equal sign for --arg=value with Slurm. This means that instead of writing --mem 96G --partition staging --immediate 5, you must use `--mem=96G --partition=staging --immediate=5.
In this respect, Slurm deviates from the GNU argument syntax where the equal sign is optional for long arguments.
"},{"location":"help/faq/#slurmstepd-says-that-hwloc_get_obj_below_by_type-fails","title":"slurmstepd says that hwloc_get_obj_below_by_type fails","text":"You can ignore the following problem:
slurmstepd: error: hwloc_get_obj_below_by_type() failing, task/affinity plugin may be required to address bug fixed in HWLOC version 1.11.5\nslurmstepd: error: task[0] unable to set taskset '0x0'\nThis is a minor failure related to Slurm and cgroups. Your job should run through successfully despite this error (that is more of a warning for end-users).
"},{"location":"help/faq/#how-can-i-share-filescollaborate-with-users-from-another-work-group","title":"How can I share files/collaborate with users from another work group?","text":"Please use projects as documented here. Projects were created for this particular purpose.
"},{"location":"help/faq/#whats-the-relation-of-charite-mdc-and-cluster-accounts","title":"What's the relation of Charite, MDC, and cluster accounts?","text":"For HPC 4 Research either an active and working Charite or MDC account is required (that is, you can login e.g., into email.charite.de or mail.mdc-berlin.de). The system has a separate meta directory that is used for the authorization of users (in other words, whether the user is active, has access to the system, and which groups the user belongs to). Charite and MDC accounts map to accounts <Charite user name>_c and <MDC user name>_m accounts in this meta directory. In the case that a user has both Charite and MDC accounts these are completely separate entities in the meta directory. For authentication (veryfing that a user has acccess to an account), the Charite and MDC account systems (MS Active Directory) are used. Authentication currently only uses the SSH keys deposited into Charite (via zugang.charite.de) and MDC (via MDC persdb). Users have to obtain a suitable Charite/MDC account via Charite and MDC central IT departments and upload their SSH keys through the host organization systems on their own. The hpc-helpdesk process is then used for getting their accounts setup on the HPC 4 Research system (the home/work/scratch shares being setup), becoming part of the special hpc-users group that controls access to the system and organizing users into work groups and projects.
The process of submitting keys to Charite and MDC is documented in the \"Connecting\" section.
"},{"location":"help/faq/#how-do-charitemdccluster-accounts-interplay-with-vpn-and-the-mdc-jail-node","title":"How do Charite/MDC/Cluster accounts interplay with VPN and the MDC jail node?","text":"Charite users have to obtain a VPN account with the appropriate VPN access permissions, i.e., Zusatzantrag B as documented here. For Charite VPN, as for all Charite IT systems, users must use their Charite user name (e.g., jdoe and not jdoe_c).
MDC users either have to use MDC VPN or the MDC jail node, as documented here. For MDC VPN and jail node, as for all MDC IT systems, users must use their MDC user name (e.g., jdoe and not jdoe_m).
For help with VPN or jail node, please contact the central Charite or MDC helpdesks as appropriate.
Only when connecting from the host organizations' VPN or from the host organizations' jail node, the users use the HPC 4 Research user name that is jdoe_c or jdoe_m and not jdoe!
BIH HPC IT does not have the resources to offer such a service to normal users.
In particular, for privacy sensitive data this comes with a large number of strings attached to fulfill all regulatory requirements. If you need to exchange such data then you need to contact the central IT departments of your home organisation:
If your data is not privacy sensitive or you can guarantee strong encryption of the data then the Gigamove service of RWTH Aachen might come in handy:
You can login via Charite/MDC credentials (or most German academic institutions) and store up to 1TB of data at a time in the account with each file having up to 100GB.
As a note, Charite GB IT has a (German) manual on how to use 7-Zip with AES256 and strong passwords for encrypting data such that it is fit for transfer over unencrypted channels. You can find it here (Charite Intranet only) at point 2.12.
The key point is using a strong password (e.g. with the pwgen utility), creating an encrypted file with AES256 encryption, using distinct password for each recipient, and exchanging the password over a second channel (SMS or voice phone). Note that the central manual remains the ground truth of information and this FAQ entry may not reflect the current process recommended by GB IT if it changes without us noticing.
Can you solve the question yourself?
Please try to solve the question yourself with this manual and Google.
If the problem turns out to be hard, we're happy to help.
This page describes how to write a good help request ticket.
There is more specific questions for common issues given below.
"},{"location":"help/good-tickets/#problems-connecting-to-the-cluster","title":"Problems Connecting to the Cluster","text":"ifconfig on Linux/Mac, ipconfig on Windows)?ssh-add -l and add -vvv to the SSH command that fails for you.scontrol show job <jobid> or sacct --long -j <jobid> of your job.Getting Help
Our helpdesk can be reached via email to hpc-helpdesk@bih-charite.de. Please read our guide on how to write good tickets first.
Please also use the handy figure below on general problem resolution.
But before contacting the helpdesk, try to get help in the HPC Talk BIH HPC user self-help forum!
"},{"location":"help/helpdesk/#helpdesk-scope","title":"Helpdesk Scope","text":"Our helpdesk can support you in the following areas:
We will try our best to resolve these issues. Please note that all other questions can only be answered in a \"best effort way\".
"},{"location":"help/helpdesk/#helpdesk-non-scope","title":"Helpdesk Non-Scope","text":"The following topics are out of scope for the BIH HPC Helpdesk:
We're happy to see if we can help when there is a concrete problem with the software, e.g.,
Another community-driven possibility to get help is our \u201cHPC Talk\u201d forum. After this manual, it should be the first place to consult.
https://hpc-talk.cubi.bihealth.org/
Its main purpose is to serve as a FAQ, so with time and more people participating, you will more likely find an answer to your question. We also use it to make announcements and give an up-to-date status of current problems with the cluster, so it is worth logging in every once in a while. It is also a great first place to look at if you're experiencing problems with the cluster. Maybe it's a known issue.
Despite users also being able to answer questions, our admins do participate on a regular basis.
"},{"location":"how-to/connect/gpu-nodes/","title":"How-To: Connect to GPU Nodes","text":"The cluster has seven nodes with four Tesla V100 GPUs each: hpc-gpu-{1..7} and one node with 10 A40 GPUs: hpc-gpu-8.
Connecting to a node with GPUs is easy. You request one or more GPU cores by adding a generic resources flag to your Slurm job submission via srun or sbatch.
--gres=gpu:tesla:COUNT will request NVIDIA V100 cores.--gres=gpu:a40:COUNT will request NVIDIA A40 cores.--gres=gpu:COUNT will request any available GPU cores.Your job will be automatically placed in the Slurm gpu partition and allocated a number of COUNT GPUs.
Info
Fair use rules apply. As GPU nodes are a limited resource, excessive use by single users is prohibited and can lead to mitigating actions. Be nice and cooperative with other users. Tip: getent passwd USER_NAME will give you a user's contact details.
Interactive Use of GPU Nodes is Discouraged
While interactive computation on the GPU nodes is convenient, it makes it very easy to forget a job after your computation is complete and let it run idle. While your job is allocated, it blocks the allocated GPUs and other users cannot use them although you might not be actually using them. Please prefer batch jobs for your GPU jobs over interactive jobs.
Furthermore, interactive GPU jobs are currently limited to 24 hours. We will monitor the situation and adjust that limit to optimize GPU usage and usability.
Please also note that allocation of GPUs through Slurm is mandatory, in other words: Using GPUs via SSH sessions is prohibited. The scheduler is not aware of manually allocated GPUs and this interferes with other users' jobs.
"},{"location":"how-to/connect/gpu-nodes/#usage-example","title":"Usage example","text":""},{"location":"how-to/connect/gpu-nodes/#preparation","title":"Preparation","text":"We will setup a miniforge installation with pytorch testing the GPU. If you already have this setup then you can skip this step
hpc-login-1:~$ srun --pty bash\nhpc-cpu-1:~$ wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-x86_64.sh\nhpc-cpu-1:~$ bash Miniforge3-Linux-x86_64.sh -b -p ~/work/miniforge\nhpc-cpu-1:~$ source ~/work/miniforge/bin/activate\nhpc-cpu-1:~$ conda create -y -n gpu-test pytorch cudatoolkit=10.2 -c pytorch\nhpc-cpu-1:~$ conda activate gpu-test\nhpc-cpu-1:~$ python -c 'import torch; print(torch.cuda.is_available())'\nFalse\nhpc-cpu-1:~$ exit\nhpc-login-1:~$\nThe False shows that CUDA is not available on the node but that is to be expected. We're only warming up!
Let us now allocate a GPU. The Slurm schedule will properly allocate GPUs for you and setup the environment variable that tell CUDA which devices are available. The following dry run shows these environment variables (and that they are not available on the login node).
hpc-login-1:~$ export | grep CUDA_VISIBLE_DEVICES\nhpc-login-1:~$ srun --gres=gpu:tesla:1 --pty bash\nhpc-gpu-1:~$ export | grep CUDA_VISIBLE_DEVICES\ndeclare -x CUDA_VISIBLE_DEVICES=\"0\"\nhpc-gpu-1:~$ exit\nhpc-login-1:~$ srun --gres=gpu:tesla:2 --pty bash\nhpc-gpu-1:~$ export | grep CUDA_VISIBLE_DEVICES\ndeclare -x CUDA_VISIBLE_DEVICES=\"0,1\"\nAs you see, you can also reserve multiple GPUs. If we were to open two concurrent connections (e. g. in a screen) to the same node when allocating one GPU each, the allocated GPUs would be non-overlapping. Note that any two jobs are isolated using Linux cgroups (\"container\" technology) so you cannot accidentally use a GPU of another job.
Now to the somewhat boring part where we show that CUDA actually works.
hpc-login-1:~$ srun --gres=gpu:tesla:1 --pty bash\nhpc-gpu-1:~$ nvcc --version\nnvcc: NVIDIA (R) Cuda compiler driver\nCopyright (c) 2005-2019 NVIDIA Corporation\nBuilt on Wed_Oct_23_19:24:38_PDT_2019\nCuda compilation tools, release 10.2, V10.2.89\nhpc-gpu-1:~$ source ~/work/miniforge/bin/activate\nhpc-gpu-1:~$ conda activate gpu-test\nhpc-gpu-1:~$ python -c 'import torch; print(torch.cuda.is_available())'\nTrue\nNote
If scheduling a GPU fails, consider explicitely requesting the GPU partion via --partition gpu (or #SBATCH --partition gpu).
Also make sure to read the FAQ entry \"I have problems connecting to the GPU node! What's wrong?\" if you encounter problems.
"},{"location":"how-to/connect/gpu-nodes/#bonus-1-who-is-using-the-gpus","title":"Bonus #1: Who is using the GPUs?","text":"Use squeue to find out about currently queued jobs (the egrep only keeps the header and entries in the gpu partition).
hpc-login-1:~$ squeue | egrep -iw 'JOBID|gpu'\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 33 gpu bash holtgrem R 2:26 1 hpc-gpu-1\nTo find out how active the GPU nodes actually are, you can connect to the nodes (without allocating a GPU; you can do this even if the node is full) and then use nvidia-smi.
hpc-login-1:~$ ssh hpc-gpu-1 bash\nhpc-gpu-1:~$ nvidia-smi\nFri Mar 6 11:10:08 2020\n+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 440.33.01 Driver Version: 440.33.01 CUDA Version: 10.2 |\n|-------------------------------+----------------------+----------------------+\n| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |\n| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |\n|===============================+======================+======================|\n| 0 Tesla V100-SXM2... Off | 00000000:18:00.0 Off | 0 |\n| N/A 62C P0 246W / 300W | 16604MiB / 32510MiB | 99% Default |\n+-------------------------------+----------------------+----------------------+\n| 1 Tesla V100-SXM2... Off | 00000000:3B:00.0 Off | 0 |\n| N/A 61C P0 270W / 300W | 16604MiB / 32510MiB | 100% Default |\n+-------------------------------+----------------------+----------------------+\n| 2 Tesla V100-SXM2... Off | 00000000:86:00.0 Off | 0 |\n| N/A 39C P0 55W / 300W | 0MiB / 32510MiB | 0% Default |\n+-------------------------------+----------------------+----------------------+\n| 3 Tesla V100-SXM2... Off | 00000000:AF:00.0 Off | 0 |\n| N/A 44C P0 60W / 300W | 0MiB / 32510MiB | 4% Default |\n+-------------------------------+----------------------+----------------------+\n\n+-----------------------------------------------------------------------------+\n| Processes: GPU Memory |\n| GPU PID Type Process name Usage |\n|=============================================================================|\n| 0 43461 C python 16593MiB |\n| 1 43373 C python 16593MiB |\n+-----------------------------------------------------------------------------+\nNote that allocating a GPU makes it unavailable for everyone else, so please behave nicely and be cooperative. If you see someone blocking the GPU nodes for a long time, first find out who it is. You can type getent passwd USER_NAME on any cluster node to see their email address (and work phone number if added). Send a friendly email, most likely they blocked the node accidentally. If you cannot resolve the issue (e. g. the user is not reachable) then please contact hpc-helpdesk@bih-charite.de.
The cluster has 4 high-memory nodes with 1.5 TB of RAM. You can connect to these nodes using the highmem SLURM partition (see below). Jobs allocating more than 200 GB of RAM are automatically routed to the highmem nodes.
Info
Fair use rules apply. As high-memory nodes are a limited resource, excessive use by single users is prohibited and can lead to mitigating actions. Be nice and cooperative with other users. Tip: getent passwd USER_NAME will give you a user's contact details.
In the cluster there are four High-memory used which can be used:
hpc-login-1:~$ sinfo -p highmem\nPARTITION AVAIL TIMELIMIT NODES STATE NODELIST \nhighmem up 14-00:00:0 3 idle med040[1-4] \nTo connect to one of them, simply allocate more than 200GB of RAM in your job.
hpc-login-1:~$ srun --pty --mem=300GB bash -i\nmed0401:~$\nYou can also pick one of the hostnames:
hpc-login-1:~$ srun --pty --mem=300GB --nodelist=med0403 bash -i\nmed0403:~$\nAfter successfull login, you can see that you are in \"highmem\" queue:
med0403:~$ squeue\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) \n[...]\n 270 highmem bash holtgrem R 1:25 1 med0403 \nClick on the edit link at the top of each page as shown below.
Please Contribute!
This guide is far from complete. Please feel free to contribute, e.g., refer to How-To: Contribute to this Document.
Please make sure that you have read How-To: Debug Software as a general primer.
As debugging is hard enough already, it makes one wonder how to do this on the HPC system in batch mode. Here is a list of pointers.
"},{"location":"how-to/misc/debug-at-hpc/#attempt-1-run-it-interactively","title":"Attempt 1: Run it interactively!","text":"First of all, you can of course get an interactive session using srun --pty bash -i and then run your program interactively. Make sure to allocate appropriate memory and cores for your purpose. You might also want to first start a screen or tmux session on the login node such that network interruptions to the login node don't harm your hard debugging work!
Does the program work correctly if you do this? If yes, and it only fails when run in batch mode, consider the following behaviour of the scheduler.
The scheduler takes your resource requirements and tries to find a free slot. Once it has found a free slot, it will attempt to run the program. This mainly differs in running it interactively in standard input, output, and error streams.
/dev/null such that no input is read. You can change this with the --input= flag to specify a file.--output=. You can use certain wildcards to make the output (but also the input files) depend on things like the job ID or job name.mkdir it in the job script itself.Please refer to the sbatch documentation for details.
If your program fails without leaving any log file or any other trace, make sure that the path to the output file exists. To the best of the author's knowledge, there is no way to tell apart a crash because this does not exist and a program failure (except maybe for the running time of 0 seconds and memory usage of 0 bytes).
"},{"location":"how-to/misc/debug-at-hpc/#attempt-2-inspect-the-logs","title":"Attempt 2: Inspect the logs","text":"Do you see any exception in your log files? If not, continue.
If your job is canceled by scancel or stopped because it exhausted it maximal running time or allocated resources then you will find a note in the last line of your error output log (usually folded into the standard output). Please note that if the previous output line did not include a line ending, the message might be at the very end of the last line.
The message will look similar to:
slurmstepd: error: *** JOB <your job id> ON med0xxx CANCELLED AT 2020-09-02T21:01:12 DUE TO TIME LIMIT ***\nIdeally, you can add one or more --verbose/-v flags to your program to increase verbosity. See how far your program gets, see where it fails. This attempt will be greatly helped by reproducible running on a minimal working example.
sattach","text":"You can use sattach for attaching your terminal to your running job. This way, you can perform an interactive inspection of the commands.
You can combine this with one of the next attempst of using debuggers to e.g., get an pdb debugger at an important position of your program. However, please note that pdb and ipdb will stop the program's execution if the standard input stream is at end of file (which /dev/null is and this is used by default in sbatch jobs).
Log into the node that your program runs on either using srun --pty --nodelist=NODE or using ssh. Please note that you should never perform computational intensive things when logging into the node directly. You can then use all activity inspection tips from How-To: Debug Software.
After having logged into the node running your program, you can of course also attach to the program with gdb -p PID or cgdb -p PID.
Here are some final remarks:
Please Contribute!
This guide is far from complete. Please feel free to contribute, e.g., refer to How-To: Contribute to this Document.
Software development in general or even debugging of software are very broad topics. As such, we will not be able to handle them here comprehensively. Rather, we will give a tour de force on practical and minimal approaches of debugging of software. Here, debugging refers to the process of locating errors in your program and removing them.
Origin of the term debugging
The terms \"bug\" and \"debugging\" are popularly attributed to Admiral Grace Hopper in the 1940s. While she was working on a Mark II computer at Harvard University, her associates discovered a moth stuck in a relay and thereby impeding operation, whereupon she remarked that they were \"debugging\" the system. However, the term \"bug\", in the sense of \"technical error\", dates back at least to 1878 and Thomas Edison (see software bug for a full discussion).
-- Wikipedia: Debugging
When forgetting a moment about everything known about software engineering, programming roughly work sin the following cycle:
You run your program. In the case of failure, you need to remove the problem until the program runs through. You then start implementing the next change or feature. But how do you actually locate the problem? Let us walk through a couple of steps.
"},{"location":"how-to/misc/debug-software/#step-1-find-out-that-there-is-an-error","title":"Step 1: Find out that there is an error","text":"This might seem trivial but let us think about this for a moment. For this
You could make this step a bit more comfortable by writing a little checker script that compares expected and actual output.
"},{"location":"how-to/misc/debug-software/#step-2-reproduce-your-error","title":"Step 2: Reproduce your error","text":"You will have to find out how often or regularly the problem occurs. Does the problem occur on all inputs or only specific ones? Does it occur with all parameters? Make sure that you can reproduce the problem, otherwise the problem will be hard to track down.
Discard randomness
In most applications, true randomness is neither required nor used in programs. Rather, pseudo random number generators are used that are usually seeded with a special value. In many cases, the current time is used which makes it hard to reproduce problems. Rather, use a fixed seed, e.g., by calling srand(42) in C. You could also make this a parameter of your program, but make sure that you can fix all pseudo randomness in your program so you can deterministically reproduce its behaviour.
Try to find a minimal input set on which you can produce your problem. For example, you could use samtools view FILE.bam chr1:90,000-100,000 to cut out regions from a BAM file. The next step is to nail down the problem. Ideally, you can deactivate or comment out parts of your program that are irrelevant to the problem.
This will allow you to get to the problematic point in your program quicker and make the whole debugging exercise easier on yourself.
"},{"location":"how-to/misc/debug-software/#interlude-what-we-have-up-to-here","title":"Interlude: What we have up to here","text":"We can now
If you reached the points above, you have probably cut the time to resolve the problem by 90% already.
Let us now consider a few things that you can do from here to find the source of your problems.
"},{"location":"how-to/misc/debug-software/#method-1-stare-at-your-source-code","title":"Method 1: Stare at your source code","text":"Again, this is trivial, but: look at your code and try to follow through what it does with your given input. This is nicely complemented with the following methods. ;-)
There is a class of tools to help you in doing this, so-called static code analysis tools. They analyze the source code for problematic patterns. The success and power of such analysis tools tends to corellate strongly with how strictly typed the targeted programming language is. E.g., there are very powerful tools for Java, C/C++. However, there is some useful tool support out there for dynamic languages such as Python.
Here is a short list of pointers to static code analysis tools (feel free to extend the list):
The most simple approach is to use print statements (or similar) to print the current line or value of parameters. While sometimes frowned upon, this certainly is one of the most robust ways to see what is happening in your program. However, beware that too much output might slow down your program or actually make your problem disappear in the case of subtle threading/timing issues (sometimes referred to as \"Heisenbugs\").
Standard output vs. error
Classically, Linux/Unix programs can print back to the user's terminal in two ways: standard output and standard errors. By convention, logging should go to stderr. The standard error stream also has the advantage that writing to it has a more direct effect. In contrast to stdout which is usually setup to be (line) buffered (you will only see output after the next newline character), stderr is unbuffered.
"},{"location":"how-to/misc/debug-software/#look-at-tophtop","title":"Look attop/htop","text":"The tools top and htop are useful tools for inspecting the activity on the current computer. The following parameters are useful (and are actually also available as key strokes when they are running).
-c -- show the programs' command lines-u USER -- show the processes of the userYou can exit either tool by pressing q or Ctrl-C.
Use the man, Luke!
Besides searching the internet for a unix command, you can also read its manual page by running man TOOL. If this does not work, try TOOL --help to see its builtin help function. Also, doing an internet search for \"man tool\" might help.
strace","text":"The program strace allows you to intercept the calls of your program to the kernel. As the kernel is needed for actions such as accessing the network or file system. Thus this is not so useful if your program gets stuck in \"user land\", but this might be useful to see which files it is accessing.
Pro-Tip: if you move the selection line of htop to a process then you can strace the program by pressing s.
lsof","text":"The lsof program lists all open files with the processes that are accessing them. This is useful for seeing which files you program has opened.
You can even build a progress bar with lsof, although that requires sudo privileges which you might not have on the system that you are using.
Pro-Tip: if you move the selection line of htop to a process then you can list the open files by pressing l.
There are more ways of inspecting your program, here are some:
perfLet us now enter the world of interactive debuggers. Integrated development environment (IDEs) generally consist of an editor, a compiler/interpreter, and an ineractive/visual debugger. Usually, they have a debugger program at their core that can also be used on their command line.
"},{"location":"how-to/misc/debug-software/#old-but-gold-gdb","title":"Old but gold:gdb","text":"On Unix systems, a widely used debugger is gdb the GNU debugger. gdb is a command line program and if you are not used to it, it might be hard to use. However, here are some pointers on how to use it:
The commands in interactive mode include:
quit or Ctrl-D to exit the debuggerb file.ext:123 set breakpoint in file.ext on line 123r run the programp var_name print the value of the variable var_namedisplay var_name print the value of the variable var_name every time execution stopsl print the source code around the current line (multiple calls will show the next 10 lines or so, and so on)l 123 print lines around line 123f show information about the current frame (that is the current source location)bt show the backtrace (that is all functions above the current one)n step to the next lines step into function callsfinish run the current function until it returnshelp to get more helpYou can call your program directly with command line arguments using cgdb [cgdb-args] --args path/to/program -- [program-args.You can also attach to running programs usingcgdb -p PIDonce you have found out the process ID to attach to usinghtoporps`.
Pro-tip: use cgdb for an easier to use version that displays the source code in split screen and stores command line histories over sessions.
You can get a simple REPL (read-execute-print loop) at virtually any position in your program by adding:
import pdb; pdb.set_trace()\nYou will get a prompt at the current position and can issue several commands including:
quit or Ctrl-D to exit the debuggerp var_name to print the variable with var_namef show information about the current frame (that is the current source location)bt show the backtrace (that is all functions called above the current one)continue to continue runninghelp to get more helpPro-tip: use import ipdb; ipdb.set_trace() (after installing the ipdb package, of course) to get an IPython-based prompt that is much more comfortable to use.
Here is a free bonus pro-tip: learn how to use version control, e.g., Git. This will allow you to go back to previous versions without problems and see current changes to your source code.
Combine the pro tip on using version control (learn Git already!) with this one: learn how to write automated tests. This will allow you to quickly narrow down problematic changes in your version control history.
Again, testing is another topic alltogether, so here are just some links to testing frameworks to get you started:
The following web resources can serve as a starting point on how to use debuggers.
We provide a user forum using the Discourse software at
First of all, visit the website for the first time: https://hpc-talk.cubi.bihealth.org
You will then be directed to our Single-Sign-On Page.
Use the appropriate button for your host organisation (MDC / Charite) where also your cluster account belongs to.
Then use the usual of your host organisation.
Clicked wrong organisation?
If you accidentally clicked the wrong institution then you need to clear your browser history up to the point where you clicked (e.g., for the last hour).
You will be shown the following screen after the first login.
You can proceed with reading the notification or split it. The site is mostly self-explanatory. let us point you at a couple of interesting things for first steps.
Here you can setup your preferences
Use the \"New Topic\" button to create a new topic. Set a meaningful title, select a suitable category (we will update the list of categories over time), and write down your question or discussion item. Finally, click \"Create Topic\" to create the new topic.
You will be directed to the page with your new topic.
You can enable email notifications to receive emails if someone answers.
"},{"location":"how-to/misc/hpc-talk/#disabling-browser-notifications","title":"Disabling Browser Notifications","text":"In your settings, you will find an option to disable browser notifications in this browser.
Or you can use the do not disturb button.
"},{"location":"how-to/misc/hpc-talk/#closing-remarks","title":"Closing Remarks","text":"We established the HPC Talk forum as a self-help forum for users. Alas, there is a number of such sites out there already that are populated by more users.
How does HPC Talk fit in?
We think it is most useful for asking questions and discussing points that are directly related to the BIH HPC system.
What alternatives do I have?
For example:
Obtaining File Boxes
At the moment, file boxes are only available to members of core facilities (e.g., genomics, bioinformatics, or metabolomics) for exchanging files for their collaboration partners. Currently, HPC users cannot use the file box mechanism on their own.
BIH HPC IT provides a file exchange server to be used by the BIH core facilities and their users. The server is located in the BIH DMZ in Buch. Users authenticate using their Charite/BIH (user@CHARITE) or MDC accounts (user@MDC-BERLIN). File exchange is organized using \"file boxes\", directories created on the server to which selected users are granted access. Access control list maintenance is done with audit-trails (\"Revisionssicherheit\") and the file access itself is also logged to comply with data protection standards.
Access from Charite Network
Access from the Charite network (IP ranges 141.x.x.x and 10.x.x.x) must connect through the Charite proxy (http://proxy.charite.de:8080). Depending on the client software that you are using, you might have to configure the proxy.
File boxes are created by the core facilities (e.g., the genomics facilities at Charite and MDC). The facility members also organize the access control. Please talk to your core facility contact on file exchange.
External users must obtain a Charite or MDC account first. Account creation is handled by the core facilities that the external user is a customer of.
"},{"location":"how-to/service/file-exchange/#file-access","title":"File Access","text":"Generally, you will be given a URL to your file box similar to https://file-exchange.bihealth.org/<file-box-id>/. The files are served over an encrypted connection using WebDAV (which uses HTTPS).
The following describes how to access the files in the box from different platforms.
"},{"location":"how-to/service/file-exchange/#from-linux","title":"From Linux","text":"We describe how to access the files on the command line using the lftp program. The program is preinstalled on the BIH (and the MDC cluster) and you should be able to just install it with yum install lftp on CentOS/Red Hat or apt-get install lftp on Ubuntu/Debian.
When using lftp, you have to add some configuration first:
# cat >>~/.lftprc <<\"EOF\"\nset ssl:verify-certificate no\nset ftp:ssl-force yes\nEOF\nIn case that you want to access the files using a graphical user interface, search Google for \"WebDAV\" and your operating system or desktop environment. File browsers such as Nautilus and Thunar have built-in WebDAV support.
"},{"location":"how-to/service/file-exchange/#connecting","title":"Connecting","text":"First, log into the machine that has lftp installed. The login nodes of the BIH cluster do not have it installed but all compute and file transfer nodes have it. Go to the data download location.
host:~$ mkdir -p ~/scratch/download_dir\nhost:~$ cd ~/scratch/download_dir\nNext, start lftp. You can open the connection using open -u <user>@<DOMAIN> https://file-exchange.bihealth.org/<file-box-id>/ (NB: there is a trailing slash) where
<user> is your user name, e.g., holtgrem,<domain> is either MDC-BERLIN or CHARITE, and<file-box-id> the file box ID from the URL provided to you.When prompted, use your normal Charite/MDC password to login.
host:download_dir$ lftp\nlftp :~> open -u holtgrem@CHARITE https://file-exchange.bihealth.org/c62910b3-c1ba-49a5-81a6-a68f1f15aef6\nPassword:\ncd ok, cwd=/c62910b3-c1ba-49a5-81a6-a68f1f15aef6\nlftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6>\nYou can find a full reference of lftp on the lftp man page. You could also use help COMMAND on the lftp prompt. For example, to look at the files of the server for a bit...
lftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6> ls\ndrwxr-xr-x -- /\ndrwxr-xr-x -- dir\n-rw-r--r-- -- file1\nlftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6> find\n./\n./dir/\n./dir/file2\n./file1\nTo download all data use mirror, e.g. with -P 4 to use four download threads.
lftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6> mirror .\nTotal: 2 directories, 3 files, 0 symlinks\nNew: 3 files, 0 symlinks\nlftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6> exit\nhost:download_dir$ tree\n.\n\u251c\u2500\u2500 dir\n\u2502\u00a0\u00a0 \u2514\u2500\u2500 file2\n\u251c\u2500\u2500 file1\n\u2514\u2500\u2500 file.txt\n\n1 directory, 3 files\nIgnoring gnutls_record_recv errors.
A common error to see is mirror: Fatal error: gnutls_record_recv: The TLS connection was non-properly terminated.. You can just ignore this.
To upload data, you can use mirror -R . which is essentially the \"reverse\" of the mirror command.
lftp holtgrem@CHARITE@file-exchange.bihealth.org:/c62910b3-c1ba-49a5-81a6-a68f1f15aef6> mirror -R\nmirror: Fatal error: gnutls_record_recv: The TLS connection was non-properly terminated.\nmirror: Fatal error: gnutls_record_recv: The TLS connection was non-properly terminated.\nmirror: Fatal error: gnutls_record_recv: The TLS connection was non-properly terminated.\nTotal: 2 directories, 3 files, 0 symlinks\nModified: 3 files, 0 symlinks\n4 errors detected\nWe recommend to use WinSCP for file transfer.
After starting WinSCP, you will see a window titled Login. Just paste the URL (e.g., https://file-exchange.bihealth.org/c62910b3-c1ba-49a5-81a6-a68f1f15aef6/) of the file box into the Host name entry field. In this case, the fields File protocol etc. will be filled automatically. Next, enter your user name as user@CHARITE or user@MDC-BERLIN (the capitalization of the part behind the @ is important). The window should now look similar to the one below.
Proxy Configuration on Charite Network
If you are on the Charite network then you have to configure the proxy. Otherwise, you have to skip this step.
Click Advanced and a window titled Advanced Site Settings will pop up. Here, select Connection / Proxy in the left side. Select HTTP for the Proxy type. Then, enter proxy.charite.de as the Proxy host name and set the Port number to 8080. The window should nwo look as below. Then, click OK to apply the proxy settings.
Finally, click Login. You can now transfer files between the file exchange server and your local computer using drag and drop between WinSCP and your local Windows File Explorer. Alternatively, you can use the two-panel view of WinSCP to transfer files as described here.
For Mac, we you can also use lftp as described above in From Linux. You can find install instructions here online.
Proxy Configuration on Charite Network
If you are on the Charite network then you must have configured the proxy appropriately. Otherwise, you have to skip this step.
You can find them in your System Preference in the Network section, in the Advanced tab of your network (e.g., WiFi).
If you want to use a graphical interface then we recommend the usage of Cyberduck. After starting the program, click Open Connection on the top left, then select WebDAV (HTTPS) and fill out the form as in the following way. Paste the file box URL into the server field and use your login name (user@CHARITE or user@MDC-BERLIN) with your usual password.
If you need to perform access through a graphical user interface on your Mac, please contact hpc-helpdesk@bihealth.org for support.
"},{"location":"how-to/service/file-exchange/#security","title":"Security","text":"The file exchange server has the fail2ban software installed and configured (Charite, MDC, and BIH IPs are excluded from this).
If you are entering your user/password incorrectly for more than 5 times in 10 minutes then your machine will be banned for one hour. This means someone else that has the same IP address from the side of the file exchange server can get you blocked. This can happen if you are in the same home or university network with NAT or if you are behind a proxy. In this case you get a \"connection refused\" error. In this case, try again in one hour.
"},{"location":"how-to/software/apptainer/","title":"Using Apptainer (with Docker Images)","text":"Note
Singularity is now Apptainer! While Apptainer provides an singularity alias for backwards compatibility, it is recommanded to adapt all workflows to use the new binary apptainer.
Apptainer (https://apptainer.org/) is a popular alternative to docker, because it does not require to run as a privileged user. Apptainer can run Docker images out-of-the-box by converting them to the apptainer image format. The following guide gives a quick dive into using docker images with apptainer.
Build on your workstation, run on the HPC
Building images using Apptainer requires root privileges. We cannot give you these permissions on the BIH HPC. Thus, you will have to build the images on your local workstation (or anywhere where you have root access). You can then run the built images on the BIH HPC.
This is also true for the --writeable flag. Apparently it needs root permissions which you don't have on the cluster.
Link ~/.apptainer to ~/work/.apptainer
Because you only have a quota of 1 GB in your home directory, you should symlink ~/.apptainer to ~/work/.apptainer.
host:~$ mkdir -p ~/work/.apptainer && ln -sr ~/work/.apptainer ~/.apptainer\nIn case you already have a apptainer directory:
host:~$ mv ~/.apptainer ~/work/.apptainer && ln -sr ~/work/.apptainer ~/.apptainer\nRun a bash in a docker image:
host:~$ apptainer shell docker://godlovedc/lolcow\nRun a command in a docker image:
host:~$ apptainer exec docker://godlovedc/lolcow echo \"hello, hello!\"\nRun a bash in a docker image, enable access to the cuda driver (--nv) and mount a path (--bind or -B):
host:~$ apptainer shell --nv --bind /path_on_host/:/path_inside_container/ docker://godlovedc/lolcow\nCaveats
Notes
APPTAINERENV_: host:~$ APPTAINERENV_HELLO=123 apptainer shell docker://godlovedc/lolcow echo $HELLO\napptainer shell or apptainer exec uses as cwd the host callers cwd not the one set in the Dockerfile. One can change this by setting --pwd.The easiest variant to run a docker image available via a docker hub is by specifying its url. This causes apptainer to download the image and convert it to a apptainer image:
host:~$ apptainer run docker://godlovedc/lolcow\nor to open a shell inside the image
host:~$ apptainer shell docker://godlovedc/lolcow\nFurthermore, similar to docker, one can pull (and convert) remote image with the following call:
host:~$ apptainer pull docker://godlovedc/lolcow\nIn case your registry requires authentication you can provide it via a prompt by adding the option --docker-login:
host:~$ apptainer pull --docker-login docker://ilumb/mylolcow\nor by setting the following environment variables:
host:~$ export APPTAINER_DOCKER_USERNAME=ilumb\nhost:~$ export APPTAINER_DOCKER_PASSWORD=<redacted>\nhost:~$ apptainer pull docker://ilumb/mylolcow\nMore details can be found in the Apptainer documentation.
"},{"location":"how-to/software/apptainer/#option-2-converting-docker-images","title":"Option 2: Converting Docker Images","text":"Another option is to convert your docker image into the Apptainer/Singularity image format. This can be easily done using the docker images provided by docker2singularity.
To convert the docker image docker_image_name to the apptainer image apptainer_image_name one can use the following command line. The output image will be located in output_directory_for_images.
host:~$ docker run -v /var/run/docker.sock:/var/run/docker.sock -v /output_directory_for_images/:/output --privileged -t --rm quay.io/singularity/docker2singularity --name apptainer_image_name docker_image_name\nThe resulting image can then directly be used as image:
host:~$ apptainer exec apptainer_image_name.sif bash\nHere are some tips for making Docker images compatible with Apptainer taken from docker2singulrity:
~/.bashrc, ~/.profile, etc.ENTRYPOINT instruction set pointing to the command line interface to your pipeline.CMD - rely only on ENTRYPOINT.ENTRYPOINT docker run -i -t --entrypoint /bin/bash bids/example.--read-only --tmpfs /run --tmpfs /tmp parameters (this emulates the read-only behavior of Apptainer).USER instruction set.from the official website: \"Cell Ranger is a set of analysis pipelines that process Chromium single-cell RNA-seq output to align reads, generate feature-barcode matrices and perform clustering and gene expression analysis\"
"},{"location":"how-to/software/cell-ranger/#installation","title":"installation","text":"requires registration before download from here
to unpack Cell Ranger, its dependencies and the cellranger script:
cd /data/cephfs-1/home/users/$USER/work\nmv /path/to/cellranger-3.0.2.tar.gz .\ntar -xzvf cellranger-3.0.2.tar.gz\nwill be provided in /data/cephfs-1/work/projects/cubit/current/static_data/app_support/cellranger
add a file slurm.template to /data/cephfs-1/home/users/$USER/work/cellranger-3.0.2/martian-cs/v3.2.0/jobmanagers/sge.template with the following contents:
#!/usr/bin/env bash\n#\n# Copyright (c) 2016 10x Genomics, Inc. All rights reserved.\n#\n# =============================================================================\n# Setup Instructions\n# =============================================================================\n#\n# 1. Add any other necessary Slurm arguments such as partition (-p) or account\n# (-A). If your system requires a walltime (-t), 24 hours (24:00:00) is\n# sufficient. We recommend you do not remove any arguments below or Martian\n# may not run properly.\n#\n# 2. Change filename of slurm.template.example to slurm.template.\n#\n# =============================================================================\n# Template\n# =============================================================================\n#\n#SBATCH -J __MRO_JOB_NAME__\n#SBATCH --export=ALL\n#SBATCH --nodes=1 --ntasks-per-node=__MRO_THREADS__\n#SBATCH --signal=2\n#SBATCH --no-requeue\n#SBATCH --partition=medium\n#SBATCH --time=24:00:00\n### Alternatively: --ntasks=1 --cpus-per-task=__MRO_THREADS__\n### Consult with your cluster administrators to find the combination that\n### works best for single-node, multi-threaded applications on your system.\n#SBATCH --mem=__MRO_MEM_GB__G\n#SBATCH -o __MRO_STDOUT__\n#SBATCH -e __MRO_STDERR__\n\n__MRO_CMD__\nnote: on newer cellranger version, slurm.template needs to go to /data/cephfs-1/home/users/$USER/work/cellranger-XX/external/martian/jobmanagers/
if that hasn't been done yet, you can use cellranger mkfastq (details to be added)
count)","text":"create a script run_cellranger.sh with these contents (consult the documentation for help:
#!/bin/bash\n\n/data/cephfs-1/home/users/$USER/work/cellranger-3.0.2/cellranger count \\\n --id=sample_id \\\n --transcriptome=/data/cephfs-1/work/projects/cubit/current/static_data/app_support/cellranger/refdata-cellranger-${species}-3.0.0\\\n --fastqs=/path/to/fastqs \\\n --sample=sample_name \\\n --expect-cells=n_cells \\\n --jobmode=slurm \\\n --maxjobs=100 \\\n --jobinterval=1000\nand then submit the job via
sbatch --ntasks=1 --mem-per-cpu=4G --time=8:00:00 -p medium -o cellranger.log run_cellranger.sh\nadd a file sge.template to /data/cephfs-1/home/users/$USER/work/cellranger-3.0.2/martian-cs/v3.2.0/jobmanagers/sge.template with the following contents:
# =============================================================================\n# Template\n# =============================================================================\n#\n#$ -N __MRO_JOB_NAME__\n#$ -V\n#$ -pe smp __MRO_THREADS__\n#$ -cwd\n#$ -P medium\n#$ -o __MRO_STDOUT__\n#$ -e __MRO_STDERR__\n#$ -l h_vmem=__MRO_MEM_GB_PER_THREAD__G\n#$ -l h_rt=08:00:00\n\n#$ -m a\n#$ -M user@email.com\n\n__MRO_CMD__\nand submit the job via
qsub -cwd -V -pe smp 1 -l h_vmem=8G -l h_rt=24:00:00 -P medium -m a -j y run_cellranger.sh\nSSH Tunnels Considered Harmful
Please use our Open OnDemand Portal for running Jupyter notebooks!
The information below is still accurate. However, many users find it tricky to get SSH tunnels working correctly. A considerable number of parts is involved and you have to get each step 100% correct. Helpdesk cannot support you in problems with SSH tunnels that are caused by incorrect usage.
"},{"location":"how-to/software/jupyter/#what-is-jupyter","title":"What is Jupyter","text":"Project Jupyter is a networking protocol for interactive computing that allows the user to write and execute code for a high number of different programming languages. The most used client is Jupyter Notebook that can be encountered in various form all over the web. Its basic principle is a document consisting of different cells, each of which contains either code (executed in place) or documentation (written in markdown). This allows one to handily describe the processed workflow.
"},{"location":"how-to/software/jupyter/#setup-and-running-jupyter-on-the-cluster","title":"Setup and running Jupyter on the cluster","text":"Install Jupyter on the cluster (via conda, by creating a custom environment)
hpc-cpu-x:~$ conda create -n jupyter jupyter\nhpc-cpu-x:~$ conda activate jupyter\n(If you want to work in a language other than python, you can install more Jupyter language kernel, see the kernel list)
Now you can start the Jupyter server session (you may want to do this in a screen & srun --pty bash -i session as jupyter keeps running while you are doing computations)
hpc-cpu-x:~$ jupyter notebook --no-browser\nCheck the port number (usually 8888) in the on output and remember it for later:
[I 23:39:40.860 NotebookApp] The Jupyter Notebook is running at:\n[I 23:39:40.860 NotebookApp] http://localhost:8888/\nBy default, Jupyter will create an access token (a link stated in the output) to protect your notebook against unauthorized access which you have to save and enter in the accessing browser. You can change this to password base authorization via jupyter notebook password. If you are running multiple server on one or more nodes, one can separate them by changing the port number by adding --port=$PORT.
This is slightly trickier as we have to create a SSH connection/tunnel with potentially multiple hops in between. The easiest way is probably to configure your .ssh/config to automatically route your connection via the login node (and possibly MDC jail). This is described in our Advanced SSH config documentation
In short,add these lines to ~/.ssh/config (replace curly parts):
Host bihcluster\n user {USER_NAME}\n HostName hpc-login-2.cubi.bihealth.org\n\nHost hpc-cpu*\n user {USER_NAME}\n ProxyJump bihcluster\nFor MDC users outside the MDC network:
Host mdcjail\n HostName ssh1.mdc-berlin.de\n User {MDC_USER_NAME}\n\nHost bihcluster\n user {USER_NAME}\n HostName hpc-login-2.cubi.bihealth.org\n\nHost hpc-cpu*\n user {USER_NAME}\n ProxyJump bihcluster\nCheck that this config is working by connecting like this: ssh hpc-cpu-1. Please note that you cannot use any resources on this node without a valid Slurm session.
Now you setup a tunnel for your running Jupyter session:
workstation:~$ ssh -N -f -L 127.0.0.1:8888:localhost:{PORT} hpc-cpu-x\n8888. The cluster node srun has sent you to determines the last argument. You should now be able to connect to your Jupyter server by typing localhost:8888 in your webbrowser (see the note about token and password above).
It can and will happen that will lose connection, either due to network problems or due to shut-down of your computer. This is not a problem at all and you will not lose data, just reconnect to your session. If your notebooks are also losing connection (you will see a colorful remark in the top right corner), reconnect and click the colorful button. If this does not work, your work is still not lost as all cells that have been executed are automatically saved anyways. Copy all unexecuted cells (those are only saved periodically) and reload the browser page (after reconnecting) with F5. (you can also open a copy of the notebook in another tab, just be aware that there may be synchronisation problems)
There are two independent steps in ending a session:
Canceling the SSH tunnel
hpc-cpu-x:~$ ps aux | grep \"$PORT\"\nThis will give you something like this:
user 54 0.0 0.0 43104 784 ? Ss 15:06 0:00 ssh -N -f -L 127.0.0.1:8888:localhost:8888 hpc-cpu-x\nuser 58 0.0 0.0 41116 1024 tty1 S 15:42 0:00 grep --color=auto 8888\nfrom which you need the process ID (here 54)
hpc-cpu-x:~$ kill -9 $PID\nShutdown the Jupyter server
Open the Jupyter session, cancel the process with {Ctrl} + {C} and confirm {y}. Make sure you saved your notebooks beforehand (though auto-save catches most things).
"},{"location":"how-to/software/jupyter/#advanced","title":"Advanced","text":"If anyone has figured out, the following might also be interesting (please add):
Because the GPU nodes med030[1-4] has four GPU units we can train a model by using multiple GPUs in parallel. This How-To gives an example with Keras 2.2.4 together and tensorflow. Finally soem hints how you can submit a job on the cluster.
Hint
With tensorflow > 2.0 and newer keras version the multi_gpu_model is deprecated and you have to use the MirroredStrategy.
we need to import the multi_gpu_model model from keras.utils and have to pass our actual model (maybe sequential Keras model) into it. In general Keras automatically configures the number of available nodes (gpus=None). This seems not to work on our system. So we have to specify the numer of GPUs, e.g. two with gpus=2. We put this in a try catch environment that it will also work on CPUs.
from keras.utils import multi_gpu_model\n\ntry: \n model = multi_gpu_model(model, gpus=2) \nexcept:\n pass\nThat's it!
Please read here on how to submit jobs to the GPU nodes.
"},{"location":"how-to/software/keras/#conda-environment","title":"Conda environment","text":"All this was tested with the following conda environment:
name: cuda channels: \n- conda-forge\n- bioconda\n- defaults\ndependencies:\n- keras=2.2.4\n- python=3.6.7\n- tensorboard=1.12.0\n- tensorflow=1.12.0\n- tensorflow-base=1.12.0\n- tensorflow-gpu=1.12.0\nNote
This information is outdated and will soon be removed.
GNU Octave as Matlab alternative
Note that GNU Octave is an Open Source alternative to Matlab. While both packages are not 100% compatible, Octave is an alternative that does not require any license management. Further, you can easily install it yourself using Conda.
Want to use the Matlab GUI?
Make sure you understand X forwarding as outline in this FAQ entry.
You can also use Open OnDemand Portal to run Matlab.
"},{"location":"how-to/software/matlab/#pre-requisites","title":"Pre-requisites","text":"You have to register with hpc-helpdesk@bih-charite.de for requesting access to the Matlab licenses. Afterwards, you can connect to the High-Memory using the license_matlab_r2016b resource (see below).
BIH has a license of Matlab R2016b for 16 seats and various licensed packages (see below). To display the available licenses:
hpc-login-1:~$ scontrol show lic\nLicenseName=matlab_r2016b\n Total=16 Used=0 Free=16 Remote=no\nMatlab is installed on all of the compute nodes:
# The following is VITAL so the scheduler allocates a license to your session.\nhpc-login-1:~$ srun -L matlab_r2016b:1 --pty bash -i\nmed0127:~$ scontrol show lic\nLicenseName=matlab_r2016b\n Total=16 Used=1 Free=15 Remote=no\nmed0127:~$ module avail\n----------------- /usr/share/Modules/modulefiles -----------------\ndot module-info null\nmodule-git modules use.own\n\n----------------------- /opt/local/modules -----------------------\ncmake/3.11.0-0 llvm/6.0.0-0 openmpi/3.1.0-0\ngcc/7.2.0-0 matlab/r2016b-0\nmed0127:~$ module load matlab/r2016b-0\nStart matlab without GUI: matlab -nosplash -nodisplay -nojvm\n Start matlab with GUI (requires X forwarding (ssh -X)): matlab\nmed0127:~$ matlab -nosplash -nodisplay -nojvm\n < M A T L A B (R) >\n Copyright 1984-2016 The MathWorks, Inc.\n R2016b (9.1.0.441655) 64-bit (glnxa64)\n September 7, 2016\n\n\nFor online documentation, see http://www.mathworks.com/support\nFor product information, visit www.mathworks.com.\n\n\n Non-Degree Granting Education License -- for use at non-degree granting, nonprofit,\n educational organizations only. Not for government, commercial, or other organizational use.\n\n>> ver\n--------------------------------------------------------------------------------------------\nMATLAB Version: 9.1.0.441655 (R2016b)\nMATLAB License Number: 1108905\nOperating System: Linux 3.10.0-862.3.2.el7.x86_64 #1 SMP Mon May 21 23:36:36 UTC 2018 x86_64\nJava Version: Java is not enabled\n--------------------------------------------------------------------------------------------\nMATLAB Version 9.1 (R2016b)\nBioinformatics Toolbox Version 4.7 (R2016b)\nGlobal Optimization Toolbox Version 3.4.1 (R2016b)\nImage Processing Toolbox Version 9.5 (R2016b)\nOptimization Toolbox Version 7.5 (R2016b)\nParallel Computing Toolbox Version 6.9 (R2016b)\nPartial Differential Equation Toolbox Version 2.3 (R2016b)\nSignal Processing Toolbox Version 7.3 (R2016b)\nSimBiology Version 5.5 (R2016b)\nStatistics and Machine Learning Toolbox Version 11.0 (R2016b)\nWavelet Toolbox Version 4.17 (R2016b)\n>> exit\nFor starting the Matlab with GUI, make sure that your client is running a X11 server and you connect with X11 forwarding enabled (e.g., ssh -X hpc-login-1.cubi.bihealth.org from the Linux command line). Then, make sure to use srun -L matlab_r2016b:1 --pty --x11 bash -i for connecting to a node with X11 forwarding enabled.
client:~$ ssh -X hpc-login-1.cubi.bihealth.org\n[...]\nhpc-login-1:~ $ srun -L matlab_r2016b:1 --pty --x11 bash -i\n[...]\nmed0203:~$ module load matlab/r2016b-0\nStart matlab without GUI: matlab -nosplash -nodisplay -nojvm\n Start matlab with GUI (requires X forwarding (ssh -X)): matlab\nmed0203:~$ matlab\n[UI will start]\nFor forcing starting in text mode can be done (as said after module load): matlab -nosplash -nodisplay -nojvm.
Also see this FAQ entry.
"},{"location":"how-to/software/matlab/#see-available-matlab-licenses","title":"See Available Matlab Licenses","text":"You can use scontrol show lic to see the currently available MATLAB license. E.g., here I am running an interactive shell in which I have requested 1 of the 16 MATLAB licenses, so 15 more remain.
$ scontrol show lic\nLicenseName=matlab_r2016b\n Total=16 Used=1 Free=15 Remote=no\nGet a checkout of our MATLAB example. Then, look around at the contents of this repository.
hpc-login-1:~$ git clone https://github.com/bihealth/bih-cluster-matlab-example.git\nhpc-login-1:~$ cd bih-cluster-matlab-example\nhpc-login-1:~$ cat job_script.sh\n#!/bin/bash\n\n# Logging goes to directory sge_log\n#SBATCH -o slurm_log/%x-%J.log\n# Keep current environment variables\n#SBATCH --export=ALL\n# Name of the script\n#SBATCH --job-name MATLAB-example\n# Allocate 4GB of RAM per core\n#SBATCH --mem 4G\n# Maximal running time of 2 hours\n#SBATCH --time 02:00:00\n# Allocate one Matlab license\n#SBATCH -L matlab_r2016b:1\n\nmodule load matlab/r2016b-0\n\nmatlab -r example\n$ cat example.m\n% Example Hello World script for Matlab.\n\ndisp('Hello world!')\ndisp('Thinking...')\n\npause(10)\n\ndisp(sprintf('The square root of 2 is = %f', sqrt(2)))\nexit\nFor submitting the script, you can do the following
hpc-login-1:~$ sbatch job_script.sh\nThis will submit a job with one Matlab license requested. If you were to submit 17 of these jobs, then at least one of them would have to wait until all licenses are free.
Matlab License Server
Note that there is a Matlab license server running on the server that will check whether 16 or less Matlab sessions are currently running. If a Matlab session is running but this is not made known to the scheduler via -L matlab_r2016b then this can lead to scripts crashing as not enough licenses are available. If this happens to you, double-check that you have specified the license requirements correctly and notify hpc-helpdesk@bih-charite.de in case of any problems. We will try to sort out the situation then.
This article describes how to build an run an OpenMPI program. We will build a simple C program that uses the OpenMPI message passing interface and run it in parallel. You should be able to go from here with other languages and more complex programs. We will use a simple Makefile for building the software.
"},{"location":"how-to/software/openmpi/#loading-openmpi-environment","title":"Loading OpenMPI Environment","text":"First, load the OpenMPI package.
hpc-login-1:~$ srun --pty bash -i\nmed0127:~$ module load openmpi/4.3.0-0\nThen, check that the installation works
med0127:~$ ompi_info | head\n Package: Open MPI root@med0127 Distribution\n Open MPI: 4.0.3\n Open MPI repo revision: v4.0.3\n Open MPI release date: Mar 03, 2020\n Open RTE: 4.0.3\n Open RTE repo revision: v4.0.3\n Open RTE release date: Mar 03, 2020\n OPAL: 4.0.3\n OPAL repo revision: v4.0.3\n OPAL release date: Mar 03, 2020\nNext, clone the OpenMPI example project from Gitlab.
med0127:~$ git clone git@github.com:bihealth/bih-cluster-openmpi-example.git\nmed0127:~$ cd bih-cluster-openmpi-example/src\nMakefile
.PHONY: default clean\n\n# configure compilers\nCC=mpicc\nCXX=mpicxx\n# configure flags\nCCFLAGS += $(shell mpicc --showme:compile)\nLDFLAGS += $(shell mpicc --showme:link)\n\ndefault: openmpi_example\n\nopenmpi_example: openmpi_example.o\n\nclean:\n rm -f openmpi_example.o openmpi_example\nopenmpi_example.c
#include <stdio.h>\n#include <mpi.h>\n\nint main(int argc, char** argv) {\n // Initialize the MPI environment\n MPI_Init(NULL, NULL);\n\n // Get the number of processes\n int world_size;\n MPI_Comm_size(MPI_COMM_WORLD, &world_size);\n\n // Get the rank of the process\n int world_rank;\n MPI_Comm_rank(MPI_COMM_WORLD, &world_rank);\n\n // Get the name of the processor\n char processor_name[MPI_MAX_PROCESSOR_NAME];\n int name_len;\n MPI_Get_processor_name(processor_name, &name_len);\n\n // Print off a hello world message\n printf(\"Hello world from processor %s, rank %d\"\n \" out of %d processors\\n\",\n processor_name, world_rank, world_size);\n\n // Finalize the MPI environment.\n MPI_Finalize();\n\n return 0;\n}\nrun_mpi.sh
#!/bin/bash\n\n# Example job script for (single-threaded) MPI programs.\n\n# Generic arguments\n\n# Job name\n#SBATCH --job-name openmpi_example\n# Maximal running time of 10 min\n#SBATCH --time 00:10:00\n# Allocate 1GB of memory per node\n#SBATCH --mem 1G\n# Write logs to directory \"slurm_log\"\n#SBATCH -o slurm_log/slurm-%x-%J.log\n\n# MPI-specific parameters\n\n# Run 64 tasks (threads/on virtual cores)\n#SBATCH --nodes 64\n\n# Make sure to source the profile.d file (not available on head nodes).\n/etc/profile.d/modules.sh\n\n# Load the OpenMPI environment module to get the runtime environment.\nmodule load openmpi/3.1.0-0\n\n# Launch the program.\nmpirun -np 64 ./openmpi_example\nThe next step is building the software
med0127:~$ make\nmpicc -c -o openmpi_example.o openmpi_example.c\nmpicc -pthread -Wl,-rpath -Wl,/opt/local/openmpi-4.0.3-0/lib -Wl,--enable-new-dtags -L/opt/local/openmpi-4.0.3-0/lib -lmpi openmpi_example.o -o openmpi_example\nmed0127:~$ ls -lh\ntotal 259K\n-rw-rw---- 1 holtgrem_c hpc-ag-cubi 287 Apr 7 23:29 Makefile\n-rwxrwx--- 1 holtgrem_c hpc-ag-cubi 8.5K Apr 8 00:15 openmpi_example\n-rw-rw---- 1 holtgrem_c hpc-ag-cubi 760 Apr 7 23:29 openmpi_example.c\n-rw-rw---- 1 holtgrem_c hpc-ag-cubi 2.1K Apr 8 00:15 openmpi_example.o\n-rwxrwx--- 1 holtgrem_c hpc-ag-cubi 1.3K Apr 7 23:29 run_hybrid.sh\n-rwxrwx--- 1 holtgrem_c hpc-ag-cubi 663 Apr 7 23:35 run_mpi.sh\ndrwxrwx--- 2 holtgrem_c hpc-ag-cubi 4.0K Apr 7 23:29 sge_log\nThe software will run outside of the MPI environment -- but in a single process only, of course.
med0127:~$ ./openmpi_example\nHello world from processor med0127, rank 0 out of 1 processors\nAll of the arguments are already in the run_mpi.sh script.
med01247:~# sbatch run_mpi.sh\nExplanation of the OpenMPI-specific arguments
--ntasks 64: run 64 processes in the MPI environment.Let's look at the slurm log file, e.g., in slurm_log/slurm-openmpi_example-3181.log.
med0124:~$ cat slurm_log/slurm-openmpi_example-*.log\nHello world from processor med0133, rank 6 out of 64 processors\nHello world from processor med0133, rank 25 out of 64 processors\nHello world from processor med0133, rank 1 out of 64 processors\nHello world from processor med0133, rank 2 out of 64 processors\nHello world from processor med0133, rank 3 out of 64 processors\nHello world from processor med0133, rank 7 out of 64 processors\nHello world from processor med0133, rank 9 out of 64 processors\nHello world from processor med0133, rank 12 out of 64 processors\nHello world from processor med0133, rank 13 out of 64 processors\nHello world from processor med0133, rank 15 out of 64 processors\nHello world from processor med0133, rank 16 out of 64 processors\nHello world from processor med0133, rank 17 out of 64 processors\nHello world from processor med0133, rank 18 out of 64 processors\nHello world from processor med0133, rank 23 out of 64 processors\nHello world from processor med0133, rank 24 out of 64 processors\nHello world from processor med0133, rank 26 out of 64 processors\nHello world from processor med0133, rank 27 out of 64 processors\nHello world from processor med0133, rank 31 out of 64 processors\nHello world from processor med0133, rank 0 out of 64 processors\nHello world from processor med0133, rank 4 out of 64 processors\nHello world from processor med0133, rank 5 out of 64 processors\nHello world from processor med0133, rank 8 out of 64 processors\nHello world from processor med0133, rank 10 out of 64 processors\nHello world from processor med0133, rank 11 out of 64 processors\nHello world from processor med0133, rank 14 out of 64 processors\nHello world from processor med0133, rank 19 out of 64 processors\nHello world from processor med0133, rank 20 out of 64 processors\nHello world from processor med0133, rank 21 out of 64 processors\nHello world from processor med0133, rank 22 out of 64 processors\nHello world from processor med0133, rank 28 out of 64 processors\nHello world from processor med0133, rank 29 out of 64 processors\nHello world from processor med0133, rank 30 out of 64 processors\nHello world from processor med0134, rank 32 out of 64 processors\nHello world from processor med0134, rank 33 out of 64 processors\nHello world from processor med0134, rank 34 out of 64 processors\nHello world from processor med0134, rank 38 out of 64 processors\nHello world from processor med0134, rank 39 out of 64 processors\nHello world from processor med0134, rank 42 out of 64 processors\nHello world from processor med0134, rank 44 out of 64 processors\nHello world from processor med0134, rank 45 out of 64 processors\nHello world from processor med0134, rank 46 out of 64 processors\nHello world from processor med0134, rank 53 out of 64 processors\nHello world from processor med0134, rank 54 out of 64 processors\nHello world from processor med0134, rank 55 out of 64 processors\nHello world from processor med0134, rank 60 out of 64 processors\nHello world from processor med0134, rank 62 out of 64 processors\nHello world from processor med0134, rank 35 out of 64 processors\nHello world from processor med0134, rank 36 out of 64 processors\nHello world from processor med0134, rank 37 out of 64 processors\nHello world from processor med0134, rank 40 out of 64 processors\nHello world from processor med0134, rank 41 out of 64 processors\nHello world from processor med0134, rank 43 out of 64 processors\nHello world from processor med0134, rank 47 out of 64 processors\nHello world from processor med0134, rank 48 out of 64 processors\nHello world from processor med0134, rank 49 out of 64 processors\nHello world from processor med0134, rank 50 out of 64 processors\nHello world from processor med0134, rank 51 out of 64 processors\nHello world from processor med0134, rank 52 out of 64 processors\nHello world from processor med0134, rank 56 out of 64 processors\nHello world from processor med0134, rank 57 out of 64 processors\nHello world from processor med0134, rank 59 out of 64 processors\nHello world from processor med0134, rank 61 out of 64 processors\nHello world from processor med0134, rank 63 out of 64 processors\nHello world from processor med0134, rank 58 out of 64 processors\nIn some cases, you want to mix multithreading (e.g., via OpenMP) with MPI to run one process with multiple threads that then can communicate via shared memory. Note that OpenMPI will let processes on the same node communicate via shared memory anyway, so this might not be necessary in all cases.
The file run_hybrid.sh shows how to run an MPI job with 8 threads each.
Note well that memory is allocated on a per-slot (thus per-thread) base!
run_hybrid.sh
#!/bin/bash\n\n# Example job script for multi-threaded MPI programs, sometimes\n# called \"hybrid\" MPI computing.\n\n# Generic arguments\n\n# Job name\n#SBATCH --job-name openmpi_example\n# Maximal running time of 10 min\n#SBATCH --time 00:10:00\n# Allocate 1GB of memory per node\n#SBATCH --mem 1G\n# Write logs to directory \"slurm_log\"\n#SBATCH -o slurm_log/slurm-%x-%J.log\n\n# MPI-specific parameters\n\n# Run 8 tasks (threads/on virtual cores)\n#SBATCH --ntasks 8\n# Allocate 4 CPUs per task (cores/threads)\n#SBATCH --cpus-per-task 4\n\n# Make sure to source the profile.d file (not available on head nodes).\nsource /etc/profile.d/modules.sh\n\n# Load the OpenMPI environment module to get the runtime environment.\nmodule load openmpi/4.0.3-0\n\n# Launch the program.\nmpirun -n 8 ./openmpi_example\nWe changed the following
Let's look at the log output:
# cat slurm_log/slurm-openmpi_example-3193.log\nHello world from processor med0133, rank 1 out of 8 processors\nHello world from processor med0133, rank 3 out of 8 processors\nHello world from processor med0133, rank 2 out of 8 processors\nHello world from processor med0133, rank 6 out of 8 processors\nHello world from processor med0133, rank 0 out of 8 processors\nHello world from processor med0133, rank 4 out of 8 processors\nHello world from processor med0133, rank 5 out of 8 processors\nHello world from processor med0133, rank 7 out of 8 processors\nEach process can now launch 4 threads (e.g., by defining export OMP_NUM_THREADS=4 before the program call).
This page gives an end-to-end example how to build and install Gromacs as an example for managing complex scientific software installs in user land. You don't have to learn or understand the specifics of Gromacs. We use it as an example as there are some actual users on the BIH cluster. However, installing it is out of scope of BIH HPC administration.
Gromacs is a good example as it is a sufficiently complex piece of software. Quite some configuration is done on the command line and there is no current software package of it in the common RPM repositories. However, it is quite well-documented and easy to install for scientific software so there is a lot to be learned.
"},{"location":"how-to/software/scientific-software/#related-documents","title":"Related Documents","text":"We will perform the following step:
Makefiles)Many scientific software packages will have more dependencies. If the dependencies are available as CentOS Core or EPEL packages (such as zlib), HPC IT administration can install them. However, otherwise you will have to install them on their own.
Warning
Do not perform the compilation on the login nodes but go to a compute node instead.
"},{"location":"how-to/software/scientific-software/#downloading-and-extracting-software","title":"Downloading and Extracting Software","text":"This is best done in your scratch directory as we don't have to keep these files around for long. Note that the files in your scratch directory will automatically be removed after 2 weeks. You can also use your work directory here.
hpc-login-1:~$ srun --pty bash -i\nmed0127:~$ mkdir $HOME/scratch/gromacs-install\nmed0127:~$ cd $HOME/scratch/gromacs-install\nmed0127:~$ wget http://ftp.gromacs.org/pub/gromacs/gromacs-2018.3.tar.gz\nmed0127:~$ tar xf gromacs-2018.3.tar.gz\nmed0127:~$ ls gromacs-2018.3\nadmin cmake COPYING CTestConfig.cmake INSTALL scripts src\nAUTHORS CMakeLists.txt CPackInit.cmake docs README share tests\nSo far so good!
"},{"location":"how-to/software/scientific-software/#perform-the-configure-step","title":"Perform the Configure Step","text":"This is the most critical step. Most scientific C/C++ software has a build step and allows for, e.g., disabling and enabling features or setting installation paths. Here, you can configure the software depending on your needs and environment. Also, it is the easiest step to mess up.
Gromac's documentation is actually quite good but the author had problems to follow it to the letter. Gromacs recommends to create an MPI and a non-MPI build but the precise way did not work. This installation creates two flavours for Gromacs 2018.3, but in a different way than the Gromacs documentation proposes.
First, here is how to configure the non-MPI flavour Gromacs wants a modern compiler, so we load gcc. We will need to note down the precise version we used so later we can load it for running Gromacs with the appropriate libraries. We will install gromacs into $HOME/work/software, which is appropriate for user-installed software, but it could also go into a group or project directory. Note that we install the software into your work directory as software installations are quite large and might go above your home quota. Also, software installations are usually not precious enough to waste resources on snapshots and backups. Also that we force Gromacs to use AVX_256 for SIMD support (Intel sandy bridge architecture) to not get unsupported CPU instruction errors.
med0127:~$ module load gcc/7.2.0-0 cmake/3.11.0-0\nmed0127:~$ module list\nCurrently Loaded Modulefiles:\n 1) gcc/7.2.0-0 2) cmake/3.11.0-0\nmed0127:~$ mkdir gromacs-2018.3-build-nompi\nmed0127:~$ cd gromacs-2018.3-build-nompi\nmed0127:~$ cmake ../gromacs-2018.3 \\\n -DGMX_BUILD_OWN_FFTW=ON \\\n -DGMX_MPI=OFF \\\n -DGMX_SIMD=AVX_256 \\\n -DCMAKE_INSTALL_PREFIX=$HOME/work/software/gromacs/2018.3\nSecond, here is how to configure the MPI flavour. Note that we are also enabling the openmpi module. We will also need the precise version here so we can later load the correct libraries. Note that we install the software into the directory gromacs-mpi but switch off shared library building as recommended by the Gromacs documentation.
med0127:~$ module load openmpi/3.1.0-0\nmed0127:~$ module list\nCurrently Loaded Modulefiles:\n 1) gcc/7.2.0-0 2) cmake/3.11.0-0 3) openmpi/4.0.3-0\nmed0127:~$ mkdir gromacs-2018.3-build-mpi\nmed0127:~$ cd gromacs-2018.3-build-mpi\nmed0127:~$ cmake ../gromacs-2018.3 \\\n -DGMX_BUILD_OWN_FFTW=ON \\\n -DGMX_MPI=ON \\\n -DGMX_SIMD=AVX_256 \\\n -DCMAKE_INSTALL_PREFIX=$HOME/work/software/gromacs-mpi/2018.3 \\\n -DCMAKE_C_COMPILER=$(which mpicc) \\\n -DCMAKE_CXX_COMPILER=$(which mpicxx) \\\n -DBUILD_SHARED_LIBS=off\nThis is simple, using -j 32 allows us to build with 32 threads. If something goes wrong: meh, the \"joys\" of compilling C software.
Getting Support for Building Software
BIH HPC IT cannot provide support for compiling scientific software. Please contact the appropriate mailing lists or forums for your scientific software. You should only contact the BIH HPC IT helpdesk only if you are sure that the problem is with the BIH HPC cluster. You should try to resolve the issue on your own and with the developers of the software that you are trying to build/use.
For the no-MPI version:
med0127:~$ cd ../cd gromacs-2018.3-build-nompi\nmed0127:~$ make -j 32\n[...]\nmed0127:~$ make install\nFor the MPI version:
med0127:~$ cd ../cd gromacs-2018.3-build-mpi\nmed0127:~$ make -j 32\n[...]\nmed0127:~$ make install\nFor Gromacs 2018.3, the following is appropriate. You should be able to use this as a template for your environment module files:
med0127:~$ mkdir -p $HOME/local/modules/gromacs\nmed0127:~$ cat >$HOME/local/modules/gromacs/2018.3 <<\"EOF\"\n#%Module\nproc ModulesHelp { } {\n puts stderr {\n Gromacs molecular simulation toolkit (non-MPI version)\n\n - http://www.gromacs.org\n }\n}\n\nmodule-whatis {Gromacs molecular simulation toolkit (non-MPI)}\n\nset root /data/cephfs-1/home/users/YOURUSER/work/software/gromacs-mpi/2018.3\n\nprereq gcc/7.2.0-0\n\nconflict gromacs\nconflict gromacs-mpi\n\nprepend-path LD_LIBRARY_PATH $root/lib64\nprepend-path LIBRARY_PATH $root/lib64\nprepend-path MANPATH $root/share/man\nprepend-path PATH $root/bin\nsetenv GMXRC $root/bin/GMXRC\nEOF\nmed0127:~$ mkdir -p $HOME/local/modules/gromacs-mpi\nmed0127:~$ cat >$HOME/local/modules/gromacs-mpi/2018.3 <<\"EOF\"\n#%Module\nproc ModulesHelp { } {\n puts stderr {\n Gromacs molecular simulation toolkit (MPI version)\n\n - http://www.gromacs.org\n }\n}\n\nmodule-whatis {Gromacs molecular simulation toolkit (MPI)}\n\nset root /data/cephfs-1/home/users/YOURUSER/work/software/gromacs-mpi/2018.3\n\nprereq openmpi/4.0.3-0\nprereq gcc/7.2.0-0\n\nconflict gromacs\nconflict gromacs-mpi\n\nprepend-path LD_LIBRARY_PATH $root/lib64\nprepend-path LIBRARY_PATH $root/lib64\nprepend-path MANPATH $root/share/man\nprepend-path PATH $root/bin\nsetenv GMXRC $root/bin/GMXRC\nEOF\nWith the next command, make your local modules files path known to the environemtn modules system.
med0127:~$ module use $HOME/local/modules\nYou can verify the result:
med0127:~$ module avail\n\n------------------ /data/cephfs-1/home/users/YOURUSER/local/modules ------------------\ngromacs/2018.3 gromacs-mpi/2018.3\n\n-------------------- /usr/share/Modules/modulefiles --------------------\ndot module-info null\nmodule-git modules use.own\n\n-------------------------- /opt/local/modules --------------------------\ncmake/3.11.0-0 llvm/6.0.0-0 openmpi/3.1.0-0\ngcc/7.2.0-0 matlab/r2016b-0 openmpi/4.0.3-0\nmodule use","text":"You can add this to your ~/.bashrc file to always execute the module use after login. Note that module is not available on the login or transfer nodes, the following should work fine:
med0127:~$ cat >>~/.bashrc <<\"EOF\"\ncase \"${HOSTNAME}\" in\n login-*|transfer-*)\n ;;\n *)\n module use $HOME/local/modules\n ;;\nesac\nEOF\nNote that the paths chosen above are sensible but arbitrary. You can install any software anywhere you have permission to -- somewhere in your user and group home, maybe a project home makes most sense on the BIH HPC, no root permissions required. You can also place the module files anywhere, as long as the module use line is appropriate.
As a best practice, you could use the following location:
$HOME/work/software as a root to install software to$HOME/work/software/$PKG/$VERSION for installing a given software package in a given version$HOME/work/software/modules as the root for modules to install$HOME/work/software/$PKG/$VERSION for the module file to load the software in a given version$HOME/work/software/modules.sh as a Bash script to contain the line module use $HOME/work/software/moduleschmod ug=rX,o= $GROUP/work/software, the upper case X is essential to only set +x on directories and not files):$GROUP/work/software as a root to install software to$GROUP/work/software/$PKG/$VERSION for installing a given software package in a given version$GROUP/work/software/modules as the root for modules to install$GROUP/work/software/$PKG/$VERSION for the module file to load the software in a given version$GROUP/work/software/modules.sh as a Bash script to contain the case Bash snippet from above but with module use $GROUP/work/software/modulesEvery time you want to use Gromacs, you can now do
med0127:~$ module load gcc/7.2.0-0 gromacs/2018.3\nor, if you want to have the MPI version:
med0127:~$ module load gcc/7.2.0-0 openmpi/4.0.3-0 gromacs-mpi/2018.3\nSomething along the lines of the following job script should be appropriate. See How-To: Build Run OpenMPI Programs for more information.
#!/bin/bash\n\n# Example job script for (single-threaded) MPI programs.\n\n# Generic arguments\n\n# Job name\n#SBATCH --job-name gromacs\n# Maximal running time of 10 min\n#SBATCH --time 00:10:00\n# Allocate 1GB of memory per CPU\n#SBATCH --mem 1G\n# Write logs to directory \"slurm_log/<name>-<job id>.log\" (dir must exist)\n#SBATCH --output slurm_log/%x-%J.log\n\n# MPI-specific parameters\n\n# Launch on 8 nodes (== 8 tasks)\n#SBATCH --ntasks 8\n# Allocate 4 CPUs per task (== per node)\n#SBATCH --cpus-per-task 4\n\n# Load the OpenMPI and GCC environment module to get the runtime environment.\nmodule load gcc/4.7.0-0\nmodule load openmpi/4.0.3-0\n\n# Make custom environment modules known. Alternative, you can \"module use\"\n# them in the session you use for submitting the job.\nmodule use $HOME/local/modules\nmodule load gromacs-mpi/2018.3\n\n# Launch the program on 8 nodes and tell Gromacs to use 4 threads for each\n# invocation.\nexport OMP_NUM_THREADS=4\nmpirun -n 8 gmx_mpi mdrun -deffnm npt_1000\nmed0127:~$ mkdir slurm_log\nmed0127:~$ sbatch job_script.sh\nSubmitted batch job 3229\nTensorFlow is a package for deep learning with optional support for GPUs. You can find the original TensorFlow installation instructions here.
This article describes how to set up TensorFlow with GPU support using Conda. This how-to assumes that you have just connected to a GPU node via srun --mem=10g --partition=gpu --gres=gpu:tesla:1 --pty bash -i (for Tesla V100 GPUs, for A400 GPUs use --gres=gpu:a40:1). Note that you will need to allocate \"enough\" memory, otherwise your python session will be Killed because of too little memory. You should read the How-To: Connect to GPU Nodes tutorial on an explanation of how to do this.
This tutorial assumes, that conda has been set up as described in [Software Management]((../../best-practice/software-installation-with-conda.md).
"},{"location":"how-to/software/tensorflow/#create-conda-environment","title":"Create conda environment","text":"We recommend that you install mamba first with conda install -y mamba and use this C++ reimplementation of the conda command as follows.
$ conda create -y -n python-tf tensorflow-gpu\n$ conda activate python-tf\nLet us verify that we have Python and TensorFlow installed. You might get different versions you could pin the version on installing with `conda create -y -n python-tf python==3.9.10 tensorflow-gpu==2.6.2
$ python --version\nPython 3.9.10\n$ python -c 'import tensorflow; print(tensorflow.__version__)'\n2.6.2\nWe thus end up with an installation of Python 3.9.10 with tensorflow 2.6.2.
"},{"location":"how-to/software/tensorflow/#run-tensorflow-example","title":"Run TensorFlow Example","text":"Let us now see whether TensorFlow has recognized our GPU correctly.
$ python\n>>> import tensorflow as tf\n>>> print(\"TensorFlow version:\", tf.__version__)\nTensorFlow version: 2.6.2\n>>> print(tf.config.list_physical_devices())\n[PhysicalDevice(name='/physical_device:CPU:0', device_type='CPU'), PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]\nYay, we can proceed to run the Quickstart Tutorial.
>>> mnist = tf.keras.datasets.mnist\n>>> (x_train, y_train), (x_test, y_test) = mnist.load_data()\n>>> x_train, x_test = x_train / 255.0, x_test / 255.0\n>>> model = tf.keras.models.Sequential([\n... tf.keras.layers.Flatten(input_shape=(28, 28)),\n... tf.keras.layers.Dense(128, activation='relu'),\n... tf.keras.layers.Dropout(0.2),\n... tf.keras.layers.Dense(10)\n... ])\n>>> predictions = model(x_train[:1]).numpy()\n>>> predictions\narray([[-0.50569224, 0.26386747, 0.43226188, 0.61226094, 0.09630793,\n 0.34400576, 0.9819117 , -0.3693726 , 0.5221357 , 0.3323232 ]],\n dtype=float32)\n>>> tf.nn.softmax(predictions).numpy()\narray([[0.04234391, 0.09141268, 0.10817807, 0.12951255, 0.07731011,\n 0.09903987, 0.18743432, 0.04852816, 0.11835073, 0.09788957]],\n dtype=float32)\n>>> loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)\n>>> loss_fn(y_train[:1], predictions).numpy()\n2.3122327\n>>> model.compile(optimizer='adam',\n... loss=loss_fn,\n... metrics=['accuracy'])\n>>> model.fit(x_train, y_train, epochs=5)\n2022-03-09 17:53:47.237997: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:185] None of the MLIR Optimization Passes are enabled (registered 2)\nEpoch 1/5\n1875/1875 [==============================] - 3s 1ms/step - loss: 0.2918 - accuracy: 0.9151\nEpoch 2/5\n1875/1875 [==============================] - 3s 1ms/step - loss: 0.1444 - accuracy: 0.9561\nEpoch 3/5\n1875/1875 [==============================] - 3s 1ms/step - loss: 0.1082 - accuracy: 0.9674\nEpoch 4/5\n1875/1875 [==============================] - 3s 1ms/step - loss: 0.0898 - accuracy: 0.9720\nEpoch 5/5\n1875/1875 [==============================] - 3s 1ms/step - loss: 0.0773 - accuracy: 0.9756\n<keras.callbacks.History object at 0x154e81360190>\n>>> model.evaluate(x_test, y_test, verbose=2)\n313/313 - 0s - loss: 0.0713 - accuracy: 0.9785\n[0.0713074803352356, 0.9785000085830688]\n>>> probability_model = tf.keras.Sequential([\n... model,\n... tf.keras.layers.Softmax()\n... ])\n>>> probability_model(x_test[:5])\n<tf.Tensor: shape=(5, 10), dtype=float32, numpy=\narray([[1.2339272e-06, 6.5599060e-10, 1.0560590e-06, 5.9356184e-06,\n 5.3691075e-12, 1.4447859e-07, 5.4218874e-13, 9.9996936e-01,\n 1.0347234e-07, 2.2147648e-05],\n [2.9887938e-06, 6.8461006e-05, 9.9991941e-01, 7.2003731e-06,\n 2.9751782e-13, 8.2818183e-08, 1.4307782e-06, 2.3203837e-13,\n 4.7433215e-07, 2.9504194e-14],\n [1.8058477e-06, 9.9928612e-01, 7.8716243e-05, 3.9140195e-06,\n 3.0842333e-05, 9.4537208e-06, 2.2774333e-05, 4.5549971e-04,\n 1.1015874e-04, 6.9138093e-07],\n [9.9978787e-01, 3.0206781e-08, 2.8528208e-05, 8.5581682e-08,\n 1.3851340e-07, 2.3634559e-06, 1.8480707e-05, 1.0153375e-04,\n 1.1583331e-07, 6.0887167e-05],\n [6.4914235e-07, 2.5808356e-08, 1.8225538e-06, 2.3215563e-09,\n 9.9588013e-01, 4.6049720e-08, 3.8903639e-07, 2.9772724e-05,\n 4.3141077e-07, 4.0867776e-03]], dtype=float32)>\n>>> exit()\nWriting Slurm jobs using TensorFlow is as easy as creating the following scripts.
tf_script.py
#/usr/bin/env python\n\nimport tensorflow as tf\nprint(\"TensorFlow version:\", tf.__version__)\nprint(tf.config.list_physical_devices())\n\nmnist = tf.keras.datasets.mnist\n\n(x_train, y_train), (x_test, y_test) = mnist.load_data()\nx_train, x_test = x_train / 255.0, x_test / 255.0\n\n\nmodel = tf.keras.models.Sequential([\n tf.keras.layers.Flatten(input_shape=(28, 28)),\n tf.keras.layers.Dense(128, activation='relu'),\n tf.keras.layers.Dropout(0.2),\n tf.keras.layers.Dense(10)\n])\n\npredictions = model(x_train[:1]).numpy()\nprint(predictions)\n\nprint(tf.nn.softmax(predictions).numpy())\n\n# ... and so on ;-)\ntf_job.sh
#!/usr/bin/bash\n\n#SBATCH --job-name=tf-job\n#SBATCH --mem=10g\n#SBATCH --partition=gpu\n#SBATCH --gres=gpu:tesla:1\n\nsource $HOME/work/miniforge/bin/activate\nconda activate python-tf\n\npython tf_script.py &>tf-out.txt\nAnd then calling
$ sbatch tf_job.sh\nYou can find the reuslts in tf-out.txt after completion.
$ cat tf-out.txt \n2022-03-09 18:05:54.628846: I tensorflow/core/platform/cpu_feature_guard.cc:142] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: SSE4.1 SSE4.2 AVX AVX2 AVX512F FMA\nTo enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.\n2022-03-09 18:05:56.999848: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1510] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 30988 MB memory: -> device: 0, name: Tesla V100-SXM2-32GB, pci bus id: 0000:18:00.0, compute capability: 7.0\nTensorFlow version: 2.6.2\n[PhysicalDevice(name='/physical_device:CPU:0', device_type='CPU'), PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]\n[[-0.07757086 0.04676083 0.9420195 -0.59902835 -0.26286742 -0.392514\n 0.3231195 -0.17169198 0.3480805 0.37013203]]\n[[0.07963609 0.09017922 0.22075593 0.04727634 0.06616627 0.05812084\n 0.11888511 0.07248258 0.12188996 0.12460768]]\nThis tutorial assumes familiarity with Linux/Unix operating systems. It also assumes that you have already connected to the cluster. We have collected some links to tutorials and manuals on the internet.
"},{"location":"hpc-tutorial/episode-0/#legend","title":"Legend","text":"Before we start with our first steps tutorial, we would like to introduce the following convention that we use throughout the series:
$ Commands are prefixed with a little dollar sign\nWhile file paths are highlighted like this: /data/cephfs-1/work/projects/cubit/current.
After connecting to the cluster, you are located on a login node. To get to your first compute node, type srun --time 7-00 --mem=8G --cpus-per-task=8 --pty bash -i which will launch an interactive Bash session on a free remote node running up to 7 days, enabling you to use 8 cores and 8 Gb memory. Typing exit will you bring back to the login node.
hpc-login-1$ srun -p long --time 7-00 --mem=8G --cpus-per-task=8 --pty bash -i\nhpc-cpu-1$ exit\n$\nSee? That was easy!
"},{"location":"hpc-tutorial/episode-0/#preparation","title":"Preparation","text":"In preparation for our first steps tutorial series, we would like you to install the software for this tutorial. In general the users on the cluster will manage their own software with the help of conda. If you haven't done so so far, please follow the instructions in installing conda first. The only premise is that you are able to log into the cluster. Make also sure that you are logged in to a computation node using srun -p medium --time 1-00 --mem=4G --cpus-per-task=1 --pty bash -i.
Now we will create a new environment, so as to not interfere with your current or planned software stack, and install into it all the software that we need during the tutorial. Run the following commands:
$ conda create -n first-steps python=3 snakemake bwa delly samtools gatk4\n$ conda activate first-steps\n(first-steps) $\nThis is part one of the \"First Steps\" BIH Cluster Tutorial. Here we will build a small pipeline with alignment and variant calling. The premise is that you have the tools installed as described in Episode 0. For this episode, please make sure that you are on a compute node. As a reminder, the command to access a compute node with the required resources is
$ srun --time 7-00 --mem=8G --cpus-per-task=8 --pty bash -i\nWe will provide you with some example FASTQ files, but you can use your own if you like. You can find the data here:
/data/cephfs-1/work/projects/cubit/tutorial/input/test_R1.fq.gz/data/cephfs-1/work/projects/cubit/tutorial/input/test_R2.fq.gzFirst, you should create a folder where the output of this tutorial will go. It would be good to have it in your work directory in /data/cephfs-1/home/users/$USER, because it is faster and there is more space available.
(first-steps) $ mkdir -p /data/cephfs-1/home/users/$USER/work/tutorial/episode1\n(first-steps) $ pushd /data/cephfs-1/home/users/$USER/work/tutorial/episode1\nQuotas / File System limits
/data/cephfs-1/home/users/$USER. The reason for this is that nightly snapshots and backups are created for this directory which are precious resources./data/cephfs-1/home/users/$USER/work. The limits are much higher here but no snapshots or backups are available./data/cephfs-1/home/users/$USER/scratch. However, files placed here are automatically removed after 2 weeks. This is only appropriate for files during download or temporary files.In general it is advisable to have a proper temporary directory available. You can create one in your ~/scratch folder and make it available to the system.
(first-steps) $ export TMPDIR=/data/cephfs-1/home/users/$USER/scratch/tmp\n(first-steps) $ mkdir -p $TMPDIR\nThe static data is located in /data/cephfs-1/work/projects/cubit/current/static_data. For our small example, the required reference genome and index can be found at:
/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta/data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fastaLet's align our data:
(first-steps) $ bwa mem -t 8 \\\n -R \"@RG\\tID:FLOWCELL.LANE\\tPL:ILLUMINA\\tLB:test\\tSM:PA01\" \\\n /data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fasta \\\n /data/cephfs-1/work/projects/cubit/tutorial/input/test_R1.fq.gz \\\n /data/cephfs-1/work/projects/cubit/tutorial/input/test_R2.fq.gz \\\n| samtools view -b \\\n| samtools sort -O BAM -T $TMPDIR -o aln.bam\n\n(first-steps) $ samtools index aln.bam\nAnd do the structural variant calling:
(first-steps) $ delly call \\\n -g /data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta \\\n aln.bam\nNote that delly will not find any variants.
"},{"location":"hpc-tutorial/episode-1/#small-variant-calling-snv-indel","title":"Small Variant Calling (SNV, indel)","text":"And now for the SNP calling (this step will take ~ 20 minutes):
(first-steps) $ gatk HaplotypeCaller \\\n -R /data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta \\\n -I aln.bam \\\n -ploidy 2 \\\n -O test.GATK.vcf\nSo this is it! We used the tools that we installed previously, accessed the reference data and ran a simple alignment and variant calling pipeline. You can access a list of all static data through this wiki, follow this link to the Static Data. You can also have a peek via:
(first-steps) $ tree -L 3 /data/cephfs-1/work/projects/cubit/current/static_data | less\nWelcome to the second episode of our tutorial series!
Once you are logged in to the cluster, you have the possibility to distribute your jobs to all the nodes that are available. But how can you do this easily? The key command to this magic is sbatch. This tutorial will show you how you can use this efficiently.
sbatch Command","text":"So what is sbatch doing for you?
You use the sbatch command in front of the script you actually want to run. sbatch then puts your job into the job queue. The job scheduler looks at the current status of the whole system and will assign the first job in the queue to a node that is free in terms of computational load. If all machines are busy, yours will wait. But your job will sooner or later get assigned to a free node.
We strongly recommend using this process for starting your computationally intensive tasks because you will get the best performance for your job and the whole system won't be disturbed by jobs that are locally blocking nodes. Thus, everybody using the cluster benefits.
You may have noticed that you run sbatch with a script, not with regular commands. The reason is that sbatch only accepts bash scripts. If you give sbatch a normal shell command or binary, it won't work. This means that we have to put the command(s) we want to use in a bash script. A skeleton script can be found at /data/cephfs-1/work/projects/cubit/tutorial/skeletons/submit_job.sh
The content of the file:
#!/bin/bash\n\n# Set a name for the job (-J or --job-name).\n#SBATCH --job-name=tutorial\n\n# Set the file to write the stdout and stderr to (if -e is not set; -o or --output).\n#SBATCH --output=logs/%x-%j.log\n\n# Set the number of cores (-c or --cpus-per-task).\n#SBATCH --cpus-per-task=8\n\n# Force allocation of the two cores on ONE node.\n#SBATCH --nodes=1\n\n# Set the total memory. Units can be given in T|G|M|K.\n#SBATCH --mem=8G\n\n# Optionally, set the partition to be used (-p or --partition).\n#SBATCH --partition=medium\n\n# Set the expected running time of your job (-t or --time).\n# Formats are MM:SS, HH:MM:SS, Days-HH, Days-HH:MM, Days-HH:MM:SS\n#SBATCH --time=30:00\n\nexport TMPDIR=/data/cephfs-1/home/users/${USER}/scratch/tmp\nmkdir -p ${TMPDIR}\nThe lines starting with #SBATCH are actually setting parameters for a sbatch command, so #SBATCH --job-name=tutorial is equal to sbatch --job-name=tutorial. Slurm will create a log file with a file name composed of the job name (%x) and the job ID (%j), e.g. logs/tutorial-XXXX.log. It will not automatically create the logs directory, we need to do this manually first. Here, we emphasize the importance of the log files! They are the first place to look if anything goes wrong.
To start now with our tutorial, create a new tutorial directory with a log directory, e.g.,
(first-steps) $ mkdir -p /data/cephfs-1/home/users/$USER/work/tutorial/episode2/logs\nand copy the wrapper script to this directory:
(first-steps) $ pushd /data/cephfs-1/home/users/$USER/work/tutorial/episode2\n(first-steps) $ cp /data/cephfs-1/work/projects/cubit/tutorial/skeletons/submit_job.sh .\n(first-steps) $ chmod u+w submit_job.sh\nNow open this file and copy the same commands we executed in the last tutorial to this file.
To keep it simple, we will put everything into one script. This is perfectly fine because the alignment and indexing are sequential. But there are two steps that could be run in parallel, namely the variant calling, because they don't depend on each other. We will learn how to do that in a later tutorial. Your file should look something like this:
#!/bin/bash\n\n# Set a name for the job (-J or --job-name).\n#SBATCH --job-name=tutorial\n\n# Set the file to write the stdout and stderr to (if -e is not set; -o or --output).\n#SBATCH --output=logs/%x-%j.log\n\n# Set the number of cores (-c or --cpus-per-task).\n#SBATCH --cpus-per-task=8\n\n# Force allocation of the two cores on ONE node.\n#SBATCH --nodes=1\n\n# Set the total memory. Units can be given in T|G|M|K.\n#SBATCH --mem=8G\n\n# Optionally, set the partition to be used (-p or --partition).\n#SBATCH --partition=medium\n\n# Set the expected running time of your job (-t or --time).\n# Formats are MM:SS, HH:MM:SS, Days-HH, Days-HH:MM, Days-HH:MM:SS\n#SBATCH --time=30:00\n\nexport TMPDIR=/data/cephfs-1/home/users/${USER}/scratch/tmp\nmkdir -p ${TMPDIR}\n\nBWAREF=/data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fasta\nREF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\nbwa mem -t 8 \\\n -R \"@RG\\tID:FLOWCELL.LANE\\tPL:ILLUMINA\\tLB:test\\tSM:PA01\" \\\n $BWAREF \\\n /data/cephfs-1/work/projects/cubit/tutorial/input/test_R1.fq.gz \\\n /data/cephfs-1/work/projects/cubit/tutorial/input/test_R2.fq.gz \\\n| samtools view -b \\\n| samtools sort -O BAM -T $TMPDIR -o aln.bam\n\nsamtools index aln.bam\n\ndelly call -g \\\n $REF \\\n aln.bam\n\ngatk HaplotypeCaller \\\n -R $REF \\\n -I aln.bam \\\n -ploidy 2 \\\n -O test.GATK.vcf\nLet's run it (make sure that you are in the tutorial/episode2 directory!):
(first-steps) $ sbatch submit_job.sh\nAnd wait for the response which will tell you that your job was submitted and which job id number it was assigned. Note that sbatch only tells you that the job has started, but nothing about finishing. You won't get any response at the terminal when the job finishes. It will take approximately 20 minutes to finish the job.
You'll probably want to see how your job is doing. You can get a list of your jobs using:
(first-steps) $ squeue --me\nNote that logins are also considered as jobs.
Identify your job by the <JOBID> (1st column) or the name of the script (3rd column). The most likely states you will see (5th column of the table):
PD pending, waiting to be submittedR runningIn the 8th column you can see that your job is very likely running on a different machine than the one you are on!
Do not use Slurm and watch or loops
The watch command is a useful tool for running commands in a loop every N seconds. For example, on your workstation you could do watch 'ping -c 3 google.com' to execute three network pings to Google every two seconds.
\ud83d\udc4e Using watch or manual loops in a cluster environment can have bad effects when querying Slurm or the shared file system. Both are shared resources and \"expensive\" queries should not be run in loops. For Slurm, this includes running squeue. The same would be true for running squeue -i which performs an internal loop.
\ud83d\udc4d Use the Slurm query commands only when you actually need the output. If you run them in an (implict or explicit) loop, then do so only for a short time and don't leave this open in a screen.
Get more information about your jobs by either passing the job id:
(first-steps) $ sstat <JOBID>\nAnd of course, watch what the logs are telling you:
(first-steps) $ tail -f logs/tutorial-<JOBID>.log\nThere will be no notification when your job is done, so it is best to watch the squeue --me command. To watch the sbatch command there is a linux command watch that you give a command to execute every few seconds. This is useful for looking for changes in the output of a command. The seconds between two executions can be set with the -n option. It is best to use -n 60 to minimize unnecessary load on the file system:
(first-steps) $ watch -n 60 squeue --me\nscancel with your job-ID: (first-steps) $ scancel <job-ID>\nThe cluster has a special way of organizing itself and by telling the cluster how long and with which priority you want your jobs to run, you can help it in this. There is a system set up on the cluster where you can enqueue your jobs to so-called partitions. partitions have different prioritites and are allowed for different running times. To get to know what partitions are available, and how to use them properly, we highly encourage you to read the cluster queues wiki page.
"},{"location":"hpc-tutorial/episode-3/","title":"First Steps: Episode 3","text":"Episode Topic 0 How can I install the tools? 1 How can I use the static data? 2 How can I distribute my jobs on the cluster (Slurm)? 3 How can I organize my jobs with Snakemake? 4 How can I combine Snakemake and Slurm?In this episode we will discuss how we can parallelize steps in a pipeline that are not dependent on each other. In the last episode we saw a case (the variant calling) that could have been potentially parallelized.
We will take care of that today. Please note that we are not going to use the sbatch command we learned earlier. Thus, this tutorial will run on the same node where you execute the script. We will introduce you to Snakemake, a tool with which we can model dependencies and run things in parallel. In the next tutorial we will learn how to submit the jobs with sbatch and Snakemake combined.
For those who know make already, Snakemake will be familiar. You can think of Snakemake being a bunch of dedicated bash scripts that you can make dependent on each other. Snakemake will start the next script when a previous one finishes, and potentially it will run things in parallel if the dependencies allow.
Snakemake can get confusing, especially if the project gets big. This tutorial will only cover the very basics of this powerful tool. For more, we highly recommend digging into the Snakemake documentation:
Every Snakemake run requires a Snakefile file. Create a new folder inside your tutorial folder and copy the skeleton:
(first-steps) $ mkdir -p /data/cephfs-1/home/users/${USER}/work/tutorial/episode3\n(first-steps) $ pushd /data/cephfs-1/home/users/${USER}/work/tutorial/episode3\n(first-steps) $ cp /data/cephfs-1/work/projects/cubit/tutorial/skeletons/Snakefile .\n(first-steps) $ chmod u+w Snakefile\nYour Snakefile should look as follows:
rule all:\n input:\n 'snps/test.vcf',\n 'structural_variants/test.vcf'\n\nrule alignment:\n input:\n '/data/cephfs-1/work/projects/cubit/tutorial/input/test_R1.fq.gz',\n '/data/cephfs-1/work/projects/cubit/tutorial/input/test_R2.fq.gz',\n output:\n bam='alignment/test.bam',\n bai='alignment/test.bam.bai',\n shell:\n r\"\"\"\n export TMPDIR=/data/cephfs-1/home/users/${{USER}}/scratch/tmp\n mkdir -p ${{TMPDIR}}\n\n BWAREF=/data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n bwa mem -t 8 \\\n -R \"@RG\\tID:FLOWCELL.LANE\\tPL:ILLUMINA\\tLB:test\\tSM:PA01\" \\\n ${{BWAREF}} \\\n {input} \\\n | samtools view -b \\\n | samtools sort -O BAM -T ${{TMPDIR}} -o {output.bam}\n\n samtools index {output.bam}\n \"\"\"\n\nrule structural_variants:\n input:\n 'alignment/test.bam'\n output:\n 'structural_variants/test.vcf'\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n delly call -o {output} -g ${{REF}} {input}\n \"\"\"\n\nrule snps:\n input:\n 'alignment/test.bam'\n output:\n 'snps/test.vcf'\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n gatk HaplotypeCaller \\\n -R ${{REF}} \\\n -I {input} \\\n -ploidy 2 \\\n -O {output}\n \"\"\"\nLet me explain. The content resembles the same steps we took in the previous tutorials. Although every step has its own rule (alignment, snp calling, structural variant calling), we could instead have written everything in one rule. It is up to you to design your rules! Note that the rule names are arbitrary and not mentioned anywhere else in the file.
But there is one primary rule: the rule all. This is the kickoff rule that makes everything run.
As you might have noticed, every rule has three main parameters: input, output and shell. input defines the files that are going into the rule, output those that are produced when executing the rule, and shell is the bash script that processes input to produce output.
Rule all does not have any output or shell, it uses input to start the chain of rules. Note that the input files of this rule are the output files of rule snps and structural_variants. The input of those rules is the output of rule alignment. This is how Snakemake processes the rules: It looks for rule all (or a rule that just has input files) and figures out how it can create the required input files with other rules by looking at their output files (the input files of one rule must be the output files of another rule). In our case it traces the workflow back to rule snps and structural_variants as they have the matching output files. They depend in return on the alignment, so the alignment rule must be executed, and this is the first thing that will be done by Snakemake.
There are also some peculiarities about Snakemake:
input or output as is done in rule alignment with the output files.input and output files in the script by writing {input} or {output}.{output.bam}${{VAR}} instead of ${VAR} but not Snakemake internal variables like {input} or {output}structural_variants we cheat a bit because delly does not produce output files if it can't find variants.touching (i.e., creating) the required output file. Snakemake has a function for doing so (call touch() on the filename).But Snakemake can do more. It is able to parse the paths of the output files and set wildcards if you want. For this your input (and output) file names have to follow a parsable scheme. In our case they do! Our FASTQ files, our only initial input files, start with test. The output of the alignment as well as the variant calling is also prefixed test. We now can modify the Snakemake file accordingly, by exchanging every occurrence of test in each input or output field with {id} (note that you could also give a different name for your variable). Only the input rule should not be touched, otherwise Snakemake would not know which value this variable should have. Your Snakefile should look now like this:
rule all:\n input:\n 'snps/test.vcf',\n 'structural_variants/test.vcf'\n\nrule alignment:\n input:\n '/data/cephfs-1/work/projects/cubit/tutorial/input/{id}_R1.fq.gz',\n '/data/cephfs-1/work/projects/cubit/tutorial/input/{id}_R2.fq.gz',\n output:\n bam='alignment/{id}.bam',\n bai='alignment/{id}.bam.bai',\n shell:\n r\"\"\"\n export TMPDIR=/data/cephfs-1/home/users/${{USER}}/scratch/tmp\n mkdir -p ${{TMPDIR}}\n\n BWAREF=/data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n bwa mem -t 8 \\\n -R \"@RG\\tID:FLOWCELL.LANE\\tPL:ILLUMINA\\tLB:test\\tSM:PA01\" \\\n ${{BWAREF}} \\\n {input} \\\n | samtools view -b \\\n | samtools sort -O BAM -T ${{TMPDIR}} -o {output.bam}\n\n samtools index {output.bam}\n \"\"\"\n\nrule structural_variants:\n input:\n 'alignment/{id}.bam'\n output:\n 'structural_variants/{id}.vcf'\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n delly call -o {output} -g ${{REF}} {input}\n \"\"\"\n\nrule snps:\n input:\n 'alignment/{id}.bam'\n output:\n 'snps/{id}.vcf'\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n gatk HaplotypeCaller \\\n -R ${{REF}} \\\n -I {input} \\\n -ploidy 2 \\\n -O {output}\n \"\"\"\nBefore we finally run this, we can make a dry run. Snakemake will show you what it would do:
(first-steps) $ snakemake -n\nIf everything looks green, you can run it for real. We provide it two cores to allow two single-threaded jobs to be run simultaneously:
(first-steps) $ snakemake -j 2\nIn the last episodes we learned about distributing a job among the cluster nodes using sbatch and how to automate and parallelize our pipeline with Snakemake. We are lucky that those two powerful commands can be combined. What is the result? You will have an automated pipeline with Snakemake that uses sbatch to distribute jobs among the cluster nodes instead of running only the same node.
The best thing is that we can reuse our Snakefile as it is and just write a wrapper script to call Snakemake. We run the script and the magic will start.
First, create a new folder for this episode:
(first-steps) $ mkdir -p /data/cephfs-1/home/users/${USER}/work/tutorial/episode4/logs\n(first-steps) $ pushd /data/cephfs-1/home/users/${USER}/work/tutorial/episode4\nAnd copy the wrapper script to this folder as well as the Snakefile (you can also reuse the one with the adjustments from the previous episode):
(first-steps) $ cp /data/cephfs-1/work/projects/cubit/tutorial/skeletons/submit_snakejob.sh .\n(first-steps) $ cp /data/cephfs-1/work/projects/cubit/tutorial/skeletons/Snakefile .\n(first-steps) $ chmod u+w submit_snakejob.sh Snakefile\nThe Snakefile is already known to you but let me explain the wrapper script submit_snakejob.sh:
#!/bin/bash\n\n# Set a name for the job (-J or --job-name).\n#SBATCH --job-name=tutorial\n\n# Set the file to write the stdout and stderr to (if -e is not set; -o or --output).\n#SBATCH --output=logs/%x-%j.log\n\n# Set the number of cores (-c or --cpus-per-task).\n#SBATCH --cpus-per-task=2\n\n# Force allocation of the two cores on ONE node.\n#SBATCH --nodes=1\n\n# Set the total memory. Units can be given in T|G|M|K.\n#SBATCH --mem=1G\n\n# Optionally, set the partition to be used (-p or --partition).\n#SBATCH --partition=medium\n\n# Set the expected running time of your job (-t or --time).\n# Formats are MM:SS, HH:MM:SS, Days-HH, Days-HH:MM, Days-HH:MM:SS\n#SBATCH --time=30:00\n\n\nexport TMPDIR=/data/cephfs-1/home/users/${USER}/scratch/tmp\nexport LOGDIR=logs/${SLURM_JOB_NAME}-${SLURM_JOB_ID}\nmkdir -p $LOGDIR\n\neval \"$($(which conda) shell.bash hook)\"\nconda activate first-steps\n\nset -x\n\nsnakemake --profile=cubi-v1 -j 2 -k -p --restart-times=2\nIn the beginning you see the #SBATCH that introduces the parameters when you provide this script to sbatch as described in the second episode. Please make sure that the logs folder exists before starting the run! We then set and export the TMPDIR and LOGDIR variables. Note that LOGDIR has a subfolder named $SLURM_JOB_NAME-$SLURM_JOB_ID that will be created for you. Snakemake will store its logfiles for this very Snakemake run in this folder. The next new thing is set -x. This simply prints to the terminal every command that is executed within the script. This is useful for debugging.
Finally, the Snakemake call takes place. With the --profile option we define that Snakemake uses the Snakemake profile at /etc/xdg/snakemake/cubi-v1. The profile will take create appropriate calls to sbatch and interpret the following settings from your Snakemake rule:
threads: the number of threads to execute the job onk, M, G, or T. You can specify EITHERresources.mem/resources.mem_mb: the memory to allocate for the whole job, ORresources.mem_per_thread: the memory to allocate for each thread.resources.time: the running time of the rule, in a syntax supported by Slurm, e.g. HH:MM:SS or D-HH:MM:SSresources.partition: the partition to submit your job into (Slurm will pick a fitting partition for you by default)resources.nodes: the number of nodes to schedule your job on (defaults to 1 and you will want to keep that value unless you want to use MPI)The other options to snakemake have the meaning:
-j 2: run at most two jobs at the same time-k: keep going even if a rule execution fails-p: print the executed shell commands--restart-times=2: restart failing jobs up to two timesIt is now time to update your Snakefile such that it actually specifies the resources mentioned above:
rule all:\n input:\n 'snps/test.vcf',\n 'structural_variants/test.vcf'\n\nrule alignment:\n input:\n '/data/cephfs-1/work/projects/cubit/tutorial/input/{id}_R1.fq.gz',\n '/data/cephfs-1/work/projects/cubit/tutorial/input/{id}_R2.fq.gz',\n output:\n bam='alignment/{id}.bam',\n bai='alignment/{id}.bam.bai',\n threads: 8\n resources:\n mem='8G',\n time='12:00:00',\n shell:\n r\"\"\"\n export TMPDIR=/data/cephfs-1/home/users/${{USER}}/scratch/tmp\n mkdir -p ${{TMPDIR}}\n\n BWAREF=/data/cephfs-1/work/projects/cubit/current/static_data/precomputed/BWA/0.7.17/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n bwa mem -t 8 \\\n -R \"@RG\\tID:FLOWCELL.LANE\\tPL:ILLUMINA\\tLB:test\\tSM:PA01\" \\\n ${{BWAREF}} \\\n {input} \\\n | samtools view -b \\\n | samtools sort -O BAM -T ${{TMPDIR}} -o {output.bam}\n\n samtools index {output.bam}\n \"\"\"\n\nrule structural_variants:\n input:\n 'alignment/{id}.bam'\n output:\n 'structural_variants/{id}.vcf'\n threads: 1\n resources:\n mem='4G',\n time='2-00:00:00',\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n delly call -o {output} -g ${{REF}} {input}\n \"\"\"\n\ndef snps_mem(wildcards, attempt):\n mem = 2 * attempt\n return '%dG' % mem\n\nrule snps:\n input:\n 'alignment/{id}.bam'\n output:\n 'snps/{id}.vcf'\n threads: 1\n resources:\n mem=snps_mem,\n time='04:00:00',\n shell:\n r\"\"\"\n REF=/data/cephfs-1/work/projects/cubit/current/static_data/reference/GRCh37/g1k_phase1/human_g1k_v37.fasta\n\n gatk HaplotypeCaller \\\n -R ${{REF}} \\\n -I {input} \\\n -ploidy 2 \\\n -O {output}\n \"\"\"\nWe thus configure the resource consumption of the rules as follows:
alignment with 8 threads and up to 8GB of memory in total with a running time of up to 12 hours,structural_variants with one thread and up to 4GB of memory in with a running time of up to 2 days,snps with one thread and running up to four hours. Instead of passing a static amount of memory, we pass a resource callable. The attempt parameter will be passed a value of 1 on the initial invocation. If variant calling with the GATK HaplotypeCaller fails then it will retry and attempt will have an incremented value on each invocation (2 on the first retry and so on). Thus, we try to do small variant calling with 2, 4, 6, and 8 GB.Finally, run the script:
(first-steps) $ sbatch submit_snakejob.sh\nIf you watch squeue --me now, you will see that the jobs are distributed to the system:
(first-steps) $ squeue --me\nPlease refer to the Snakemake documentation for more details on using Snakemake, in particular how to use the cluster configuration on how to specify the resource requirements on a per-rule base.
"},{"location":"misc/external-resources/","title":"External Resources","text":""},{"location":"misc/external-resources/#basic-linux","title":"Basic Linux","text":"The BIH HPC uses CentOS Linux. A basic understanding of Linux is required. Even better, you should already have intermediate to advanced Linux/Unix skills.
BIH HPC IT cannot provide you with basic Unix training. Please ask your home organization (e.g., Charite or MDC) to provide you with basic Linux training.
That said, here are some resources that we find useful:
"},{"location":"misc/external-resources/#internet-tutorials","title":"Internet Tutorials","text":"There is a large number of Linux tutorials online including:
GOBLET has a number of Bioinformatics-focused tutorials. This includes
Some software is provided by HPC Administration based on the criteria that it is:
Currently, this includes:
On the GPU node, this also includes a recent NVIDIA CUDA version.
To see which software is available, use module avail on a compute node (this will not work on login nodes):
$ module avail\n--------------------- /opt/local/modules ---------------------\ncmake/3.11.0-0 llvm/6.0.0-0\ngcc/7.2.0-0 openmpi/4.0.3-0\nTo load software, use module load. This will adjust the environment variables accordingly, in particular update PATH such that the executable are available.
$ which gcc\n/bin/gcc\n$ module load gcc/7.2.0-0\n$ which gcc\n/opt/local/gcc-7.2.0-0/bin/gcc\nProblems with executing module?
See the corresponding FAQ entry in the case that you get a -bash: module: command not found when calling module.
The BIH Cluster is a valuable resource. It has been used to support the publications listed below.
Chopra A, Song J, Weiner J, Beule D, Schaefer AS (2024). \u201cGenetic analysis of cis-enhancers associated with bone mineral density and periodontitis in the gene SOST.\u201d ResearchSquare. \u261e Link
Opitz B, Essex M, Pascual-Leone BM, Loeber U, Kuhring M, Zhang B, Bruening U, Fritsche-Guenther R, Krzanowski M, Vernengo FF, Brumhard S, R\u00f6wekamp I, Bielecka A, Lesker T, Wyler E, Landthaler M, Mantei A, Meisel C, Caesar S, Thibeault C, Corman V, Marko L, Suttorp N, Strowig T, Kurth F, Sander L, Li Y, Kirwan J, Forslund S (2023). \u201cGut microbiota dysbiosis is associated with altered tryptophan metabolism and dysregulated inflammatory response in severe COVID-19.\u201d doi:10.21203/rs.3.rs-2459463/v1 \u261e Link
Zuljan E, von der Emde B, Piwonski I, Pestana ACA, Klinghammer K, Mock A, Horak P, Heining C, Klauschen F, Pretzell M, Boerries I, Brandts C, Kreutzfeldt S, Teleanu M, H\u00fcbschmann D, Morris L, Keller U, Glimm H, Fr\u00f6hling S, Ochsenreither S, Keilholz U, Blanc E, Beule D, Rieke D (2024). \u201cIntegrative analysis reveals a macrophage-predominant, immunosuppressive immune microenvironment and subtype-specific therapeutic vulnerabilities in advanced salivary gland cancer.\u201d medRxiv, 2024-06. doi:10.1101/2024.06.11.24308538 \u261e Link
Yerinde C, Keye J, Durlanik S, Freise I, Nowak F, Hsiao H, Letizia M, Schlickeiser S, Obermayer B, Huck A, Friedrich M, Wu H, Kunkel D, K\u00fchl AA, Bauer S, Thiel A, Siegmund B, Glauben R, Weidinger C (2022). \u201cHDAC7 controls anti-viral and anti-tumor immunity by CD8+ T cells.\u201d bioRxiv. doi:10.1101/2022.09.18.508452 \u261e Link
Winkler W, D\u00edaz C, Blanc E, Napieczynska H (2021). \u201cPreclinical models of human multiple myeloma subgroups.\u201d bioRxiv. doi:10.1101/2021.08.28.458010 \u261e Link
G\u00f6ppel T, Obermayer B, Chen I, Gerland U (2021). \u201cA kinetic error filtering mechanism for enzyme-free copying of nucleic acid sequences.\u201d bioRxiv. doi:10.1101/2021.08.06.455386 \u261e Link
"},{"location":"misc/publication-list/#2024","title":"2024","text":"K\u00f6hler RM, Binns TS, Merk T, Zhu G, Yin Z, Zhao B, Chikermane M, Vanhoecke J, Busch JL, Habets JGV, Faust K, Schneider GH, Cavallo A, Haufe S, Zhang J, K\u00fchn AA, Haynes JD, Neumann WJ. Dopamine and deep brain stimulation accelerate the neural dynamics of volitional action in Parkinson\u2019s disease. Brain. 2024 Oct 1;147(10):3358\u201369. doi: 10.1093/brain/awae219
Binns TS, K\u00f6hler RM, Vanhoecke J, Chikermane M, Gerster M, Merk T, Pellegrini F, Busch JL, Habets JGV, Cavallo A, Beyer JC, Al-Fatly B, Li N, Horn A, Krause P, Faust K, Schneider GH, Haufe S, K\u00fchn AA, Neumann WJ. Shared pathway-specific network mechanisms of dopamine and deep brain stimulation for the treatment of Parkinson\u2019s disease. bioRxiv. 2024; p. 2024.04.14.586969. Available from: doi: 10.1101/2024.04.14.586969v2
Hollunder, B., Ostrem, J.L., Sahin, I.A., Rajamani, N., Oxenford, S., Butenko, K., Neudorfer, C., Reinhardt, P., Zvarova, P., Polosan, M., Akram, H., Vissani, M., Zhang, C., Sun, B., Navratil, P., Reich, M.M., Volkmann, J., Yeh, F.-C., Baldermann, J.C., Dembek, T.A., Visser-Vandewalle, V., Alho, E.J.L., Franceschini, P.R., Nanda, P., Finke, C., K\u00fchn, A.A., Dougherty, D.D., Richardson, R.M., Bergman, H., DeLong, M.R., Mazzoni, A., Romito, L.M., Tyagi, H., Zrinzo, L., Joyce, E.M., Chabardes, S., Starr, P.A., Li, N., Horn, A., 2024. Mapping dysfunctional circuits in the frontal cortex using deep brain stimulation. Nat. Neurosci. 1\u201314. doi: 10.1038/s41593-024-01570-1
Schmidt A, Danyel M, Grundmann K, Brunet T, Klinkhammer H, Hsieh T, Engels H, Peters S, Knaus A, Moosa S, Averdunk L, Boschann F, Sczakiel HL, Schwartzmann S, Mensah MA, Pantel JT, Holtgrewe M, B\u00f6sch A, Wei\u00df C, Weinhold N, Suter A, Stoltenburg C, Neugebauer J, Kallinich T, Kaindl AM, Holzhauer S, B\u00fchrer C, Bufler P, Kornak U, Ott C, Sch\u00fclke M, Nguyen HHP, Hoffjan S, Grasemann C, Rothoeft T, Brinkmann F, Matar N, Sivalingam S, Perne C, Mangold E, Kreiss M, Cremer K, Betz RC, M\u00fccke M, Grigull L, Klockgether T, Spier I, Heimbach A, Bender T, Brand F, Stieber C, Morawiec AM, Karakostas P, Sch\u00e4fer VS, Bernsen S, Weydt P, Castro-Gomez S, Aziz A, Grobe-Einsler M, Kimmich O, Kobeleva X, \u00d6nder D, Lesmann H, Kumar S, Tacik P, Basin MA, Incardona P, Lee-Kirsch MA, Berner R, Schuetz C, K\u00f6rholz J, Kretschmer T, Donato ND, Schr\u00f6ck E, Heinen A, Reuner U, Han\u00dfke A, Kaiser FJ, Manka E, Munteanu M, Kuechler A, Cordula K, Hirtz R, Schlapakow E, Schlein C, Lisfeld J, Kubisch C, Herget T, Hempel M, Weiler-Normann C, Ullrich K, Schramm C, Rudolph C, Rillig F, Groffmann M, Muntau A, Tibelius A, Schwaibold EMC, Schaaf CP, Zawada M, Kaufmann L, Hinderhofer K, Okun PM, Kotzaeridou U, Hoffmann GF, Choukair D, Bettendorf M, Spielmann M, Ripke A, Pauly M, M\u00fcnchau A, Lohmann K, H\u00fcning I, Hanker B, B\u00e4umer T, Herzog R, Hellenbroich Y, Westphal DS, Strom T, Kovacs R, Riedhammer KM, Mayerhanser K, Graf E, Brugger M, Hoefele J, Oexle K, Mirza-Schreiber N, Berutti R, Schatz U, Krenn M, Makowski C, Weigand H, Schr\u00f6der S, Rohlfs M, Vill K, Hauck F, Borggraefe I, M\u00fcller-Felber W, Kurth I, Elbracht M, Knopp C, Begemann M, Kraft F, Lemke JR, Hentschel J, Platzer K, Strehlow V, Jamra RA, Kehrer M, Demidov G, Beck-W\u00f6dl S, Graessner H, Sturm M, Zeltner L, Sch\u00f6ls LJ, Magg J, Bevot A, Kehrer C, Kaiser N, Turro E, Horn D, Gr\u00fcters-Kieslich A, Klein C, Mundlos S, N\u00f6then M, Riess O, Meitinger T, Krude H, Krawitz PM, Haack T, Ehmke N, Wagner M (2024). \u201cNext-generation phenotyping integrated in a national framework for patients with ultrarare disorders improves genetic diagnostics and yields new molecular findings.\u201d Nature Genetics. ISSN 1546-1718, doi:10.1038/s41588-024-01836-1 \u261e Link
Marot-Lassauzaie V, Beneyto-Calabuig S, Obermayer B, Velten L, Beule D, Haghverdi L (2024). \u201cIdentifying cancer cells from calling single-nucleotide variants in scRNA-seq data.\u201d Bioinformatics, 40(9), btae512. ISSN 1367-4811, doi:10.1093/bioinformatics/btae512 \u261e Link
Horn V, Cancino CA, Steinheuer LM, Obermayer B, Fritz K, Nguyen AL, Juhran KS, Plattner C, B\u00f6sel D, Oldenburg L, Burns M, Schulz AR, Saliutina M, Mantzivi E, Lissner D, Conrad T, Mashreghi M, Zundler S, Sonnenberg E, Schumann M, Haag L, Beule D, Flatz L, Atreya I, Atreya R, Bacher P, Becker C, Bojarski C, Britzen-Laurent N, Bosch-Voskens C, Chang H, Diefenbach A, G\u00fcnther C, Hegazy AN, Hildner K, Klose CS, Koop K, Krug S, K\u00fchl AA, Leppkes M, L\u00f3pez-Posadas R, Ludwig LS, Neufert C, Neurath M, Patankar J, Pr\u00fc\u00df M, Radbruch A, Romagnani C, Ronchi F, Sanders A, Scheffold A, Schulzke J, Schumann M, Sch\u00fcrmann S, Siegmund B, St\u00fcrzl M, Trajanoski Z, Triantafyllopoulou A, Waldner M, Weidinger C, Wirtz S, Zundler S, Trjanoski Z, D\u2019Haens G, Weidinger C, Mei HE, Siegmund B, Thurley K, Hegazy AN (2024). \u201cMultimodal Profiling of Peripheral Blood Identifies Proliferating Circulating Effector CD4+ T Cells as Predictors for Response to Integrin \u03b14\u03b27\u2013Blocking Therapy in Inflammatory Bowel Disease.\u201d Gastroenterology. ISSN 0016-5085, doi:10.1053/j.gastro.2024.09.021 \u261e Link
Menzel M, Martis-Thiele M, Goldschmid H, Ott A, Romanovsky E, Siemanowski-Hrach J, Seillier L, Br\u00fcchle NO, Maurer A, Lehmann K, Begemann M, Elbracht M, Meyer R, Dintner S, Claus R, Meier-Kolthoff JP, Blanc E, M\u00f6bs M, Joosten M, Benary M, Basitta P, H\u00f6lscher F, Tischler V, Gro\u00df T, Kutz O, Prause R, William D, Horny K, Goering W, Sivalingam S, Borkhardt A, Blank C, Junk SV, Yasin L, Moskalev EA, Carta MG, Ferrazzi F, T\u00f6gel L, Wolter S, Adam E, Matysiak U, Rosenthal T, D\u00f6nitz J, Lehmann U, Schmidt G, Bartels S, Hofmann W, Hirsch S, Dikow N, G\u00f6bel K, Banan R, Hamelmann S, Fink A, Ball M, Neumann O, Rehker J, Kloth M, Murtagh J, Hartmann N, Jurmeister P, Mock A, Kumbrink J, Jung A, Mayr E, Jacob A, Trautmann M, Kirmse S, Falkenberg K, Ruckert C, Hirsch D, Immel A, Dietmaier W, Haack T, Marienfeld R, F\u00fcrstberger A, Niew\u00f6hner J, Gerstenmaier U, Eberhardt T, Greif PA, Appenzeller S, Maurus K, Doll J, Jelting Y, Jonigk D, M\u00e4rkl B, Beule D, Horst D, Wulf A, Aust D, Werner M, Reuter-Jessen K, Str\u00f6bel P, Auber B, Sahm F, Merkelbach-Bruse S, Siebolts U, Roth W, Lassmann S, Klauschen F, Gaisa NT, Weichert W, Evert M, Armeanu-Ebinger S, Ossowski S, Schroeder C, Schaaf CP, Malek N, Schirmacher P, Kazdal D, Pfarr N, Budczies J, Stenzinger A (2024). \u201cBenchmarking whole exome sequencing in the German network for personalized medicine.\u201d European Journal of Cancer, 211, 114306. ISSN 0959-8049, doi:10.1016/j.ejca.2024.114306 \u261e Link
Fatykhova D, Fritsch VN, Siebert K, Methling K, Lalk M, Busche T, Kalinowski J, Weiner J, Beule D, Bertrams W, others (2024). \u201cMicroenvironmental acidification by pneumococcal sugar consumption fosters barrier disruption and immune suppression in the human alveolus.\u201d European Respiratory Journal. doi:doi.org/10.1183/13993003.01983-2023 \u261e Link
Adhikary PP, Idowu T, Tan Z, Hoang C, Shanta S, Dumbani M, Mappalakayil L, Awasthi B, Bermudez M, Weiner J, Beule D, Wolber G, Page BD, Hedtrich S (2024). \u201cDisrupting TSLP-TSLP receptor interactions via putative small molecule inhibitors yields a novel and efficient treatment option for atopic diseases.\u201d EMBO Molecular Medicine, 1-27. doi:10.1038/s44321-024-00085-3 \u261e Link
Mattei D, Ivanov A, Hammer J, Ugursu B, Schalbetter S, Richetto J, Weber-Stadlbauer U, Mueller F, Scarborough J, Wolf SA, Kettenmann H, Wollscheid B, Beule D, Meyer U (2024). \u201cMicroglia undergo molecular and functional adaptations to dark and light phases in male laboratory mice.\u201d Brain, Behavior, and Immunity, 120, 571-583. ISSN 0889-1591, doi:10.1016/j.bbi.2024.07.007 \u261e Link
Kopp J, Koch LA, Lyubenova H, K\u00fcchler O, Holtgrewe M, Ivanov A, Dubourg C, Launay E, Brachs S, Mundlos S, Pellerin L, Mameli M, Paolicelli RC (2024). \u201cLoss-of-function variants affecting the STAGA complex component SUPT7L cause a developmental disorder with generalized lipodystrophy.\u201d Human Genetics, 143(5), 683-694. doi:10.1007/s00439-024-02669-y \u261e Link
Ruiz-Orera J, Miller DC, Greiner J, Genehr C, Grammatikaki A, Blachut S, Mbebi J, Patone G, Myronova A, Adami E, others (2024). \u201cEvolution of translational control and the emergence of genes and open reading frames in human and non-human primate hearts.\u201d Nature Cardiovascular Research, 1-19. doi:10.1038/s44161-024-00544-7 \u261e Link
Fisch A, Pestana A, Sachse V, Doll C, Hofmann E, Heiland M, Obermueller T, Heidemann J, Dommerich S, Schoppe D, Schallenberg S, Piwonski I, Blanc E, Tinhofer I (2024). \u201cFeasibility analysis of using patient-derived tumour organoids for treatment decision guidance in locally advanced head and neck squamous cell carcinoma.\u201d European journal of cancer (Oxford, England : 1990), 213, 115100. ISSN 1879-0852, doi:10.1016/j.ejca.2024.115100 \u261e Link
Al-Diab O, S\u00fcnkel C, Blanc E, Catar RA, Ashraf MI, Zhao H, Wang P, Rinschen MM, Fritsche-Guenther R, Grahammer F, Bachmann S, Beule D, Kirwan JA, Rajewsky N, Huber TB, G\u00fcrgen D, Kusch A (2024). \u201cSex-specific molecular signature of mouse podocytes in homeostasis and in response to pharmacological challenge with rapamycin.\u201d Biology of sex differences, 15, 72. ISSN 2042-6410, doi:10.1186/s13293-024-00647-7 \u261e Link
Trinks A, Milek M, Beule D, Kluge J, Florian S, Sers C, Horst D, Morkel M, Bischoff P (2024). \u201cRobust detection of clinically relevant features in single-cell RNA profiles of patient-matched fresh and formalin-fixed paraffin-embedded (FFPE) lung cancer tissue.\u201d Cellular Oncology, 47, 1-11. doi:10.5281/zenodo.7852154 \u261e Link
\u0160mid A, \u0160tajdohar M, Milek M, Urban\u010di\u010d D, Ku\u017eeli\u010dki NK, Tamm R, Metspalu A, Mlinari\u010d-Ra\u0161\u010dan I (2024). \u201cTranscriptome analysis reveals involvement of thiopurine S-methyltransferase in oxidation-reduction processes.\u201d European Journal of Pharmaceutical Sciences, 192, 106616. doi:10.1016/j.ejps.2023.106616 \u261e Link
D\u00f6rner PJ, Anandakumar H, R\u00f6wekamp I, Vernengo FF, Pascual-Leone BM, Krzanowski M, Sellmaier J, Br\u00fcning U, Fritsche-Guenther R, Pfannkuch L, Kurth F, Milek M, Igbokwe V, L\u00f6ber U, Gutbier B, Holstein M, Heinz GA, Mashreghi M, Schulte LN, Klatt A, Caesar S, Wienhold S, Offermanns S, Mack M, Witzenrath M, Jordan S, Beule D, Kirwan JA, Forslund SK, Wilck N, Bartolomaeus H, Heimesaat MM, Opitz B (2024). \u201cClinically used broad-spectrum antibiotics compromise inflammatory monocyte-dependent antibacterial defense in the lung.\u201d Nature Communications, 15(1), 2788.
R\u00f6wekamp I, Maschirow L, Rabes A, Vernengo FF, Hamann L, Heinz GA, Mashreghi M, Caesar S, Milek M, Fonseca ACF, Wienhold S, Nouailles G, Yao L, Mousavi S, Bruder D, Boehme JD, Puzianowska-Kuznicka M, Beule D, Witzenrath M, Group CS, L\u00f6hning M, Klose CSN, Heimesaat MM, Diefenbach A, Opitz B, Fuchs A, Engelmann M, Paul G, Ayoub M, Groehl K, Riedl K, Stolz D, Bauer W, Diehl-Wiesenecker EC, Terue IvW, Galtung N, Suttorp N, Witzenrath M, Wildberg C, Pley C, Zessin E, Schmager S, Schaaf B, Kremling J, Nickoleit-Bitzenberger D, Azzaui H, Hower M, Hempel F, Prebeg K, Popkirova K, Kolditz M, Schulte-Hubbert B, Langner S, Rohde G, Bellinghausen C, Gr\u03cbnewaldt A, Endres A, Frigerio C, Fiedler B, Panning M, Welte T, Pink I, Drick N, F\u03cbhner T, Klooster Mv, Steinberg T, Barten-Neiner G, Kr\u00f6ner W, Unruh O, Adaskina N, Eberhardt F, Julius C, Illig T, Klopp N, Pletz M, Schleenvoigt BT, Bahrs C, Moeser A, Ankert J, Sommerwerck U, Wintermantel T, Dr\u00f6mann D, Parschke P, Franzen K, Rupp J, Waldeck F, K\u00e4ding N, Spinner C, Erber J, Voit F, Schneider J, Falcone M, Tiseo G, Heigener D, Hering I, Albrich W, Rassouli F, Wirth B, Neurohr C, Essig A, Stenger S, Wallner M, Burgmann H, Traby L, Schubert L, nurses aas (2024). \u201cIL-33 controls IL-22-dependent antibacterial defense by modulating the microbiota.\u201d Proceedings of the National Academy of Sciences, 121(22), e2310864121. doi:10.1073/pnas.2310864121 \u261e Link
Wei T, Blanc E, Peidli S, Bischoff P, Trinks A, Horst D, Sers C, Bl\u00fcthgen N, Beule D, Morkel M, Obermayer B (2024). \u201cHigh-confidence calling of normal epithelial cells allows identification of a novel stem-like cell state in the colorectal cancer microenvironment.\u201d International Journal of Cancer, n/a(n/a). doi:10.1002/ijc.35079 \u261e Link
Hramyka D, Sczakiel HL, Zhao MX, Stolpe O, Nieminen M, Adam R, Danyel M, Einicke L, H\u00e4gerling R, Knaus A, Mundlos S, Schwartzmann S, Seelow D, Ehmke N, Mensah MA, Boschann F, Beule D, Holtgrewe M (2024). \u201cREEV: review, evaluate and explain variants.\u201d Nucleic Acids Research, gkae366. doi:10.1093/nar/gkae366 \u261e Link
Ramberger E, Sapozhnikova V, Ng YLD, Dolnik A, Ziehm M, Popp O, Str\"ang E, Kull M, Gr\"unschl\"ager F, Kr\"uger J, Benary M, M\"uller S, Gao X, Murgai A, Haji M, Schmidt A, Lutz R, Nogai A, Braune J, Laue D, Langer C, Khandanpour C, Bassermann F, D\"ohner H, Engelhardt M, Straka C, Hundemer M, Beule D, Haas S, Keller U, Einsele H, Bullinger L, Knop S, Mertins P, Kr\"onke J (2024). \u201cThe proteogenomic landscape of multiple myeloma reveals insights into disease biology and therapeutic opportunities.\u201d Nature Cancer. ISSN 2662-1347, doi:10.1038/s43018-024-00784-3 \u261e Link
Habringer S, Ihlow J, Kleo K, Klostermann A, Schmidt M, Chai L, Kn\u00f6dler M, Leyvraz S, Sigler C, Sinn B, Maschmeyer G, Jegodka Y, Benary M, Ott C, Tinhofer I, Sch\u00e4fer R, M\u00f6bs M, Keller U, Keilholz U, Rieke DT (2024). \u201cA diagnostic challenge of KIT p.V559D and BRAF p.G469A mutations in a paragastric mass.\u201d The Oncologist, oyae137. ISSN 1083-7159, doi:10.1093/oncolo/oyae137 \u261e Link
Shi W, Song J, Weiner J, Chopra A, Dommisch H, Beule D, Schaefer AS (2024). \u201clncRNA CDKN2B-AS1 regulates collagen expression.\u201d Human Genetics. doi:10.1007/s00439-024-02674-1 \u261e Link
Shi W, Song J, Dommisch H, Weiner J, Schaefer A (2024). \u201cSTAT1 binds to CAD and periodontitis risk SNP rs10757278, causing inhibition of lncRNA CDKN2B-AS1 expression, which activates transcription of collagen genes in cells of the extracellular matrix.\u201d In EUROPEAN JOURNAL OF HUMAN GENETICS, volume 32, 402-402. SPRINGERNATURE CAMPUS, 4 CRINAN ST, LONDON, N1 9XW, ENGLAND.
Zheng L, Chopra A, Weiner J, Dommisch H, Schaefer A (2024). \u201cSystematic search for microRNA regulated genes with roles in the etiology of periodontitis.\u201d In EUROPEAN JOURNAL OF HUMAN GENETICS, volume 32, 406-406. SPRINGERNATURE CAMPUS, 4 CRINAN ST, LONDON, N1 9XW, ENGLAND.
Lebedin M, Ratswohl C, Garg A, Schips M, Garc\u00eda CV, Spatt L, Thibeault C, Obermayer B, Weiner J, Vel\u00e1squez IM, others (2024). \u201cSoluble ACE2 correlates with severe COVID-19 and can impair antibody responses.\u201d iScience.
Tranter E, Frentsch M, H\u00fctter-Kr\u00f6nke ML, Vuong GL, Busch D, Loyal L, Henze L, Rosnev S, Blau I, Thiel A, Beule D, Bullinger L, Obermayer B, Na I (2024). \u201cComparable CD8+ T-cell responses to SARS-CoV-2 vaccination in single-cell transcriptomics of recently allogeneic transplanted patients and healthy individuals.\u201d Journal of Medical Virology, 96(3), e29539. doi:10.1002/jmv.29539 \u261e Link
Bhasin MA, Knaus A, Incardona P, Schmid A, Holtgrewe M, Elbracht M, Krawitz PM, Hsieh T (2024). \u201cEnhancing Variant Prioritization in VarFish through On-Premise Computational Facial Analysis.\u201d Genes, 15(3), 370.
Ermshaus A, Piechotta M, R\u00fcter G, Keilholz U, Leser U, Benary M (2024). \u201cpreon: Fast and accurate entity normalization for drug names and cancer types in precision oncology.\u201d Bioinformatics, 40(3), btae085. ISSN 1367-4811, doi:10.1093/bioinformatics/btae085 \u261e Link
Jahn D, Knapstein PR, Otto E, K\u00f6hli P, Sevecke J, Graef F, Graffmann C, Fuchs M, Jiang S, Rickert M, Erdmann C, Appelt J, Revend L, K\u00fcttner Q, Witte J, Rahmani A, Duda G, Xie W, Donat A, Schinke T, Ivanov A, Tchouto MN, Beule D, Frosch K, Baranowsky A, Tsitsilonis S, Keller J (2024). \u201cIncreased beta2-adrenergic signaling promotes fracture healing through callus neovascularization in mice.\u201d Science Translational Medicine, 16(743), eadk9129. doi:10.1126/scitranslmed.adk9129 \u261e Link
"},{"location":"misc/publication-list/#2023","title":"2023","text":"Monsorno K, Ginggen K, Ivanov A, Buckinx A, Lalive AL, Tchenio A, Benson S, Vendrell M, D\u2019Alessandro A, Beule D, others (2023). \u201cLoss of microglial MCT4 leads to defective synaptic pruning and anxiety-like behavior in mice.\u201d Nature Communications, 14(1), 5749. doi:10.1038/s41467-023-41502-4 \u261e Link
Markus D, Pelletier A, Boube M, Port F, Boutros M, Payre F, Obermayer B, Zanet J (2023). \u201cThe pleiotropic functions of Pri smORF peptides synchronize leg development regulators.\u201d PLOS Genetics, 19(10), 1-27. doi:10.1371/journal.pgen.1011004 \u261e Link
Ludwik KA, Opitz R, Jyrch S, Megges M, Weiner J, Beule D, K\u00fchnen P, Stachelscheid H (2023). \u201cGeneration of iPSC lines with SLC16A2: G410R or SLC16A2 knock out.\u201d Stem Cell Research, 103256. doi:10.1016/j.scr.2023.103256 \u261e Link
Schneemann M, Heils L, Moos V, Wei\u00df F, Krug SM, Weiner J, Beule D, Gerhard R, Schulzke J, B\u00fccker R (2023). \u201cA Colonic Organoid Model Challenged with the Large Toxins of Clostridioides difficile TcdA and TcdB Exhibit Deregulated Tight Junction Proteins.\u201d Toxins, 15(11), 643. doi:10.3390/toxins15110643 \u261e Link
Monsorno K, Ginggen K, Ivanov A, Buckinx A, Lalive AL, Tchenio A, Benson S, Vendrell M, D\u2019Alessandro A, Beule D, Pellerin L, Mameli M, Paolicelli RC (2023). \u201cLoss of microglial MCT4 leads to defective synaptic pruning and anxiety-like behavior in mice.\u201d Nature Communications, 14(1), 5749.
Zheng L, Chopra A, Weiner J, Beule D, Dommisch H, Schaefer A (2023). \u201cmiRNAs from inflamed gingiva link gene signaling to increased MET expression.\u201d Journal of Dental Research, 00220345231197984. doi:10.1177/002203452311979 \u261e Link
Mhlekude B, Postmus D, Stenzel S, Weiner J, Jansen J, Zapatero-Belinch\u00f2n FJ, Olmer R, Richter A, Heinze J, Heinemann N, M\u00fchlemann B, Schroeder S, Jones TC, M\u00fcller MA, Drosten C, Pich A, Thiel V, Martin U, Niemeyer D, Gerold G, Beule D, Goffinet C (2023). \u201cPharmacological inhibition of bromodomain and extra-terminal proteins induces an NRF-2-mediated antiviral state that is subverted by SARS-CoV-2 infection.\u201d PLoS pathogens, 19(9), e1011657. doi:10.1371/journal.ppat.1011657 \u261e Link
Benary M, Wang XD, Schmidt M, Soll D, Hilfenhaus G, Nassir M, Sigler C, Kn\u00f6dler M, Keller U, Beule D, Keilholz U, Leser U, Rieke DT (2023). \u201cLeveraging Large Language Models for Decision Support in Personalized Oncology.\u201d JAMA Network Open, 6(11), e2343689-e2343689. doi:10.1001/jamanetworkopen.2023.43689
Briest F, Noerenberg D, Hennch C, Yoshida K, Hablesreiter R, Nimo J, Sasca D, Kirchner M, Mansouri L, Inoue Y, Wiegand L, Staiger AM, Casadei B, Korkolopoulou P, Weiner J, Lopez-Guillermo A, Warth A, Schneider T, Nagy \u00c1, Klapper W, Hummel M, Kanellis G, Anagnostopoulos I, Mertins P, Bullinger L, Rosenquist R, Vassilakopoulos TP, Ott G, Ogawa S, Damm F (2023). \u201cFrequent ZNF217 mutations lead to transcriptional deregulation of interferon signal transduction via altered chromatin accessibility in B cell lymphoma.\u201d Leukemia, 1-13. doi:10.1038/s41375-023-02013-9 \u261e Link
Spranger L, Weiner J, Bredow J, Zeitz U, Grittner U, Boschmann M, Dickmann S, Stob\u00e4us N, Schwartzenberg RJ, Brachs M, Spranger J, Mai K (2023). \u201cThrifty energy phenotype predicts weight regain in postmenopausal women with overweight or obesity and is related to FGFR1 signaling.\u201d Clinical Nutrition, 42(4), 559-567. doi:10.1016/j.clnu.2023.02.020 \u261e Link
Chopra A, Song J, Weiner J, Keceli HG, Dincer PR, Cruz R, Carracedo A, Blanco J, Dommisch H, Schaefer AS (2023). \u201cRSPO4 is a potential risk gene of stages III-IV, grade C periodontitis through effects on innate immune response and oral barrier integrity.\u201d Journal of Clinical Periodontology, 50(4), 476-486. doi:10.1111/jcpe.13758 \u261e Link
Lechner L, Opitz R, Silver MJ, Krabusch PM, Prentice AM, Field MS, Stachelscheid H, Leit\u00e3o E, Schr\u00f6der C, Vallone VF, Horsthemke B, J\u00f6ckel K, Schmidt B, N\u00f6then MM, Hoffmann P, Herms S, Kleyn PW, Megges M, Blume-Peytavi U, Weiss K, Mai K, Blankenstein O, Obermayer B, Wiegand S, K\u00fchnen PKP (2023). \u201cEarly-set POMC methylation variability is accompanied by increased risk for obesity and is addressable by MC4R agonist treatment.\u201d Science Translational Medicine, 15(705), eadg1659. doi:10.1126/scitranslmed.adg1659 \u261e Link
Obermayer B, Keilholz L, Conrad T, Frentsch M, Blau I, Vuong L, Lesch S, Movasshagi K, Tietze-Stolley C, Loyal L, Henze L, Penack O, Stervbo U, Babel N, Haas S, Beule D, Bullinger L, Wittenbecher F, Na I (2023). \u201cSingle-cell clonal tracking of persistent T-cells in allogeneic hematopoietic stem cell transplantation.\u201d Frontiers in Immunology, 14. ISSN 1664-3224, doi:10.3389/fimmu.2023.1114368 \u261e Link
Gressier E, Schulte-Schrepping J, Petrov L, Brumhard S, Stubbemann P, Hiller A, Obermayer B, Spitzer J, Kostevc T, Whitney PG, Bachem A, Odainic A, van de Sandt C, Nguyen THO, Ashhurst T, Wilson K, Oates CVL, Gearing LJ, Meischel T, Hochheiser K, Greyer M, Clarke M, Kreutzenbeck M, Gabriel SS, Kastenm\u00fcller W, Kurts C, Londrigan SL, Kallies A, Kedzierska K, Hertzog PJ, Latz E, Chen YE, Radford KJ, Chopin M, Schroeder J, Kurth F, Gebhardt T, Sander LE, Sawitzki B, Schultze JL, Schmidt SV, Bedoui S (2023). \u201cCD4+ T cell calibration of antigen-presenting cells optimizes antiviral CD8+ T cell immunity.\u201d Nature Immunology, 24(6), 979-990. ISSN 1529-2916, doi:10.1038/s41590-023-01517-x \u261e Link
Loedige I, Baranovskii A, Mendonsa S, Dantsuji S, Popitsch N, Breimann L, Zerna N, Cherepanov V, Milek M, Ameres S, Chekulaeva M (2023). \u201cmRNA stability and m6A are major determinants of subcellular mRNA localization in neurons.\u201d Molecular Cell, 83(15), 2709-2725.e10. ISSN 1097-2765, doi:10.1016/j.molcel.2023.06.021 \u261e Link
Boeddrich A, Haenig C, Neuendorf N, Blanc E, Ivanov A, Kirchner M, Schleumann P, Bayraktaroglu I, Richter M, Molenda CM, Sporbert A, Zenkner M, Schnoegl S, Suenkel C, Schneider L, Rybak-Wolf A, Kochnowsky B, Byrne LM, Wild EJ, Nielsen JE, Dittmar G, Peters O, Beule D, Wanker EE (2023). \u201cA proteomics analysis of 5xFAD mouse brain regions reveals the lysosome-associated protein Arl8b as a candidate biomarker for Alzheimer's disease.\u201d Genome Med, 15, 50. doi:10.1186/s13073-023-01206-2 \u261e Link
de Bortoli T, Benary M, Horak P, Lamping M, Stintzing S, Tinhofer I, Leyvraz S, Sch\u00e4fer R, Klauschen F, Keller U, Rieke DT, Jelas I (2023). \u201cTumour mutational burden and survival with molecularly matched therapy.\u201d European Journal of Cancer, 190, 112925. doi:10.1016/j.ejca.2023.05.013 \u261e Link
Stellmach C, Sass J, Auber B, Boeker M, Wienker T, Heidel AJ, Benary M, Schumacher S, Ossowski S, Klauschen F, M\u00f6ller Y, Schmutzler R, Ustjanzew A, Werner P, Tomczak A, H\u00f6lter T, Thun S (2023). \u201cCreation of a structured molecular genomics report for Germany as a local adaption of HL7\u2019s Genomic Reporting Implementation Guide.\u201d Journal of the American Medical Informatics Association. ISSN 1527-974X, doi:10.1093/jamia/ocad061 \u261e Link
Winkler W, D\u00edaz CF, Blanc E, Napieczynska H, Langner P, Werner M, Walter B, Wollert-Wulf B, Yasuda T, Heuser A, Beule D, Mathas S, Anagnostopoulos I, Rosenwald A, Rajewsky K, Janz M (2023). \u201cMouse models of human multiple myeloma subgroups.\u201d Proceedings of the National Academy of Sciences of the United States of America, 120, e2219439120. ISSN 1091-6490, doi:10.1073/pnas.2219439120 \u261e Link
Spranger L, Weiner J, Bredow J, Zeitz U, Grittner U, Boschmann M, Dickmann S, Stobaus N, Schwartzenberg JJ, Brachs M, Spranger J, Mai K (2023). \u201cThrifty energy phenotype predicts weight regain in postmenopausal women with overweight or obesity and is related to FGFR1 signaling.\u201d Clinical Nutrition, 42, 559-567. doi:10.1016/j.clnu.2023.02.020 \u261e Link
Freitag K, Eede P, Ivanov A, Sterczyk N, Schneeberger S, Borodina T, Sauer S, Beule D, Heppner F (2023). \u201cDiverse but unique astrocytic phenotypes during embryonic stem cell differentiation, culturing and development.\u201d Commun Biol, 6, 40. doi:10.1038/s42003-023-04410-3 \u261e Link
Nieminen M, Stolpe O, Kuhring M, Weiner J, Pett P, Beule D, Holtgrewe M (2023). \u201cSODAR: managing multiomics study data and metadata.\u201d Gigascience, 12, giad052. doi:10.1093/gigascience/giad052 \u261e Link
Graff P, Wilzopolski J, Voss A, Blimkie TM, Weiner J3, Kershaw O, Panwar P, Hackett T, Br\u00f6mme D, Loyal L, Thiel A, Beule D, Hancock RE, Gruber AD, B\u00e4umer W, Hedtrich S (2023). \u201cExtracellular matrix remodeling in atopic dermatitis harnesses the onset of an asthmatic phenotype and is a potential contributor to the atopic march.\u201d Journal of Investigative Dermatology. doi:10.1016/j.jid.2023.09.278
M\u00fcller R, Freitag-Wolf S, Weiner J3, Chopra A, Top T, Dommisch H, Schaefer A (2023). \u201cCase-only design identifies interactions of genetic risk variants at SIGLEC5 and PLG with the lncRNA CTD-2353F22.1 implying the importance of periodontal wound healing for disease aetiology.\u201d J Clin Periodontol, 50, 90-101. doi:10.1111/jcpe.13712 \u261e Link
Mothes R, Pascual-Reguant A, Koehler R, Liebeskind J, Liebheit A, Bauherr S, Philipsen L, Dittmayer C, Laue M, von Manitius R, Elezkurtaj S, Durek P, Heinrich F, Heinz GA, Guerra GM, Obermayer B, Meinhardt J, Ihlow J, Radke J, Heppner FL, Enghard P, Stockmann H, Aschman T, Schneider J, Corman VM, Sander LE, Mashreghi M, Conrad T, Hocke AC, Niesner RA, Radbruch H, Hauser AE (2023). \u201cDistinct tissue niches direct lung immunopathology via CCL18 and CCL21 in severe COVID-19.\u201d Nature Communications, 14(1), 791. ISSN 2041-1723, doi:10.1038/s41467-023-36333-2 \u261e Link
Rieke DT, Schr\u00f6der S, Schafhausen P, Blanc E, Zuljan E, von der Emde B, Beule D, Keller U, Keilholz U, Klinghammer K (2023). \u201cTargeted treatment in a case series of AR+, HRAS/PIK3CA co-mutation salivary duct carcinoma.\u201d Frontiers in Oncology, 13. ISSN 2234-943X, doi:10.3389/fonc.2023.1107134
"},{"location":"misc/publication-list/#2022","title":"2022","text":"Holle J, Bartolomaeus H, L\u00f6ber U, Behrens F, Bartolomaeus TUP, Anandakumar H, Wimmer MI, Vu DL, Kuhring M, Br\u00fcning U, Maifeld A, Geisberger S, Kempa S, Schumacher F, Kleuser B, Bufler P, Querfeld U, Kitschke S, Engler D, Kuhrt LD, Drechsel O, Eckardt K, Forslund SK, Th\u00fcrmer A, McParland V, Kirwan JA, Wilck N, M\u00fcller D (2022). \u201cInflammation in Children with CKD Linked to Gut Dysbiosis and Metabolite Imbalance.\u201d Journal of the American Society of Nephrology, 33(12), 2259. ISSN 1046-6673, doi:10.1681/ASN.2022030378 \u261e Link
Baumgardt M, H\u00fclsemann M, L\u00f6wa A, Fatykhova D, Hoffmann K, Kessler M, Mieth M, Hellwig K, Frey D, Langenhagen A, Voss A, Obermayer B, Wyler E, D\u00f6kel S, Gruber AD, T\u00f6lch U, Hippenstiel S, Hocke AC, H\u00f6nzke K (2022). \u201cState-of-the-art analytical methods of viral infections in human lung organoids.\u201d PLOS ONE, 17(12), 1-8. doi:10.1371/journal.pone.0276115 \u261e Link
Pennitz P, Kirsten H, Friedrich VD, Wyler E, Goekeri C, Obermayer B, Heinz GA, Mashreghi M, B\"uttner M, Trimpert J, Landthaler M, Suttorp N, Hocke AC, Hippenstiel S, T\"onnies M, Scholz M, Kuebler WM, Witzenrath M, Hoenzke K, Nouailles G (2022). \u201cA pulmonologist's guide to perform and analyse cross-species single lung cell transcriptomics.\u201d European Respiratory Review, 31(165). ISSN 0905-9180, doi:10.1183/16000617.0056-2022 \u261e Link
Bertrams W, H\u00f6nzke K, Obermayer B, T\u00f6nnies M, Bauer TT, Schneider P, Neudecker J, R\u00fcckert JC, Stiewe T, Nist A, Eggeling S, Suttorp N, Wolff T, Hippenstiel S, Schmeck B, Hocke AC (2022). \u201cTranscriptomic comparison of primary human lung cells with lung tissue samples and the human A549 lung cell line highlights cell type specific responses during infections with influenza A virus.\u201d Scientific Reports, 12(1), 20608. ISSN 2045-2322, doi:10.1038/s41598-022-24792-4 \u261e Link
Jarick K, Topczewska P, Jakob M, Yano H, Arifuzzaman M, Gao X, Boulekou S, Stokic-Trtica V, Lecl\u00e8re P, Preu\u00dfer A, Rompe Z, Stamm A, Tsou A, Chu C, Heinrich F, Guerra G, Durek P, Ivanov A, Beule D, Helfrich S, Duerr C, K\u00fchl A, Stehle C, Romagnani C, Mashreghi M, Diefenbach A, Artis D, Klose C (2022). \u201cNon-redundant functions of group 2 innate lymphoid cells.\u201d Nature, 611(7937), 794-800. doi:10.1038/s41586-022-05395-5 \u261e Link
Ivanov A, Mattei D, Radscheit K, Compagnion A, Pett J, Herzel H, Paolicelli R, Piwecka M, Meyer U, Beule D (2022). \u201cAnalyses of circRNA Expression throughout the Light-Dark Cycle Reveal a Strong Regulation of Cdr1as, Associated with Light Entrainment in the SCN.\u201d Int J Mol Sci, 23. doi:10.3390/ijms232012347 \u261e Link
Wirges A, Bunse M, Joedicke J, Blanc E, Gudipati V, Moles M, Shiku H, Beule D, Huppa J, H\u00f6pken U, Rehm A (2022). \u201cEBAG9 silencing exerts an immune checkpoint function without aggravating adverse effects.\u201d Mol Ther, 30, 3358-3378. doi:10.1016/j.ymthe.2022.07.009 \u261e Link
H\u00f6nzke K, Obermayer B, Mache C, Fathykova D, Kessler M, D\u00f6kel S, Wyler E, Baumgardt M, L\u00f6wa A, Hoffmann K, Graff P, Schulze J, Mieth M, Hellwig K, Demir Z, Biere B, Brunotte L, Mecate-Zambrano A, Bushe J, Dohmen M, Hinze C, Elezkurtaj S, T\u00f6nnies M, Bauer T, Eggeling S, Tran H, Schneider P, Neudecker J, R\u00fcckert J, Schmidt-Ott K, Busch J, Klauschen F, Horst D, Radbruch H, Radke J, Heppner F, Corman V, Niemeyer D, M\u00fcller M, Goffinet C, Mothes R, Pascual-Reguant A, Hauser A, Beule D, Landthaler M, Ludwig S, Suttorp N, Witzenrath M, Gruber A, Drosten C, Sander L, Wolff T, Hippenstiel S, Hocke A (2022). \u201cHuman lungs show limited permissiveness for SARS-CoV-2 due to scarce ACE2 levels but virus-induced expansion of inflammatory macrophages.\u201d Eur Respir J, 60. doi:10.1183/13993003.02725-2021 \u261e Link
Freitag K, Sterczyk N, Wendlinger S, Obermayer B, Schulz J, Farztdinov V, M\u00fclleder M, Ralser M, Houtman J, Fleck L, Braeuning C, Sansevrino R, Hoffmann C, Milovanovic D, Sigrist S, Conrad T, Beule D, Heppner F, Jendrach M (2022). \u201cSpermidine reduces neuroinflammation and soluble amyloid beta in an Alzheimer's disease mouse model.\u201d J Neuroinflammation, 19, 172. doi:10.1186/s12974-022-02534-7 \u261e Link
Schalbetter S, von Arx A, Cruz-Ochoa N, Dawson K, Ivanov A, Mueller F, Lin H, Amport R, Mildenberger W, Mattei D, Beule D, F\u00f6ldy C, Greter M, Notter T, Meyer U (2022). \u201cAdolescence is a sensitive period for prefrontal microglia to act on cognitive development.\u201d Sci Adv, 8, eabi6672. doi:10.1126/sciadv.abi6672 \u261e Link
Grunert C, Willimsky G, Peuker C, Rhein S, Hansmann L, Blankenstein T, Blanc E, Beule D, Keller U, Pezzutto A, Busse A (2022). \u201cIsolation of Neoantigen-Specific Human T Cell Receptors from Different Human and Murine Repertoires.\u201d Cancers (Basel), 14. doi:10.3390/cancers14071842 \u261e Link
Dannebaum R, Suwalski P, Asgharian H, Du ZG, Lin H, Weiner J, Holtgrewe M, Thibeault C, M\u00fcller M, Wang X, Karadeniz Z, Saccomanno J, Doehn J, H\u00fcbner R, Hinzmann B, Bl\u00fcher A, Siemann S, Telman D, Suttorp N, Witzenrath M, Hippenstiel S, Skurk C, Poller W, Sander L, Beule D, Kurth F, Guettouche T, Landmesser U, Berka J, Luong K, Pa-COVID SG, Rubelt F, Heidecker B (2022). \u201cHighly multiplexed immune repertoire sequencing links multiple lymphocyte classes with severity of response to COVID-19.\u201d EClinicalMedicine, 48, 101438. doi:10.1016/j.eclinm.2022.101438 \u261e Link
Horak P, Griffith M, Danos A, Pitel B, Madhavan S, Liu X, Chow C, Williams H, Carmody L, Barrow-Laing L, Rieke D, Kreutzfeldt S, Stenzinger A, Tamborero D, Benary M, Rajagopal P, Ida C, Lesmana H, Satgunaseelan L, Merker J, Tolstorukov M, Campregher P, Warner J, Rao S, Natesan M, Shen H, Venstrom J, Roy S, Tao K, Kanagal-Shamanna R, Xu X, Ritter D, Pagel K, Krysiak K, Dubuc A, Akkari Y, Li X, Lee J, King I, Raca G, Wagner A, Li M, Plon S, Kulkarni S, Griffith O, Chakravarty D, Sonkin D (2022). \u201cStandards for the classification of pathogenicity of somatic variants in cancer (oncogenicity): Joint recommendations of Clinical Genome Resource (ClinGen), Cancer Genomics Consortium (CGC), and Variant Interpretation for Cancer Consortium (VICC).\u201d Genet Med, 24, 986-998. doi:10.1016/j.gim.2022.01.001 \u261e Link
Weiner J, Obermayer B, Beule D (2022). \u201cVenn Diagrams May Indicate Erroneous Statistical Reasoning in Transcriptomics.\u201d Front Genet, 13, 818683. doi:10.3389/fgene.2022.818683 \u261e Link
Gloaguen Y, Kirwan J, Beule D (2022). \u201cDeep Learning-Assisted Peak Curation for Large-Scale LC-MS Metabolomics.\u201d Anal Chem, 94, 4930-4937. doi:10.1021/acs.analchem.1c02220 \u261e Link
Chopra A, Mueller R, Weiner J3, Rosowski J, Dommisch H, Grohmann E, Schaefer A (2022). \u201cBACH1 Binding Links the Genetic Risk for Severe Periodontitis with ST8SIA1.\u201d J Dent Res, 101, 93-101. doi:10.1177/00220345211017510 \u261e Link
Bischoff P, Trinks A, Wiederspahn J, Obermayer B, Pett J, Jurmeister P, Elsner A, Dziodzio T, R\u00fcckert J, Neudecker J, Falk C, Beule D, Sers C, Morkel M, Horst D, Klauschen F, Bl\u00fcthgen N (2022). \u201cThe single-cell transcriptional landscape of lung carcinoid tumors.\u201d Int J Cancer, 150, 2058-2071. doi:10.1002/ijc.33995 \u261e Link
Georg P, Astaburuaga-Garc\u00eda R, Bonaguro L, Brumhard S, Michalick L, Lippert L, Kostevc T, G\u00e4bel C, Schneider M, Streitz M, Demichev V, Gem\u00fcnd I, Barone M, Tober-Lau P, Helbig E, Hillus D, Petrov L, Stein J, Dey H, Paclik D, Iwert C, M\u00fclleder M, Aulakh S, Djudjaj S, B\u00fclow R, Mei H, Schulz A, Thiel A, Hippenstiel S, Saliba A, Eils R, Lehmann I, Mall M, Stricker S, R\u00f6hmel J, Corman V, Beule D, Wyler E, Landthaler M, Obermayer B, von Stillfried S, Boor P, Demir M, Wesselmann H, Suttorp N, Uhrig A, M\u00fcller-Redetzky H, Nattermann J, Kuebler W, Meisel C, Ralser M, Schultze J, Aschenbrenner A, Thibeault C, Kurth F, Sander L, Bl\u00fcthgen N, Sawitzki B, PA-COVID-19 SG (2022). \u201cComplement activation induces excessive T cell cytotoxicity in severe COVID-19.\u201d Cell, 185(3), 493-512.e25. ISSN 0092-8674, doi:10.1016/j.cell.2021.12.040 \u261e Link
Zinnall U, Milek M, Minia I, Vieira-Vieira C, M\u00fcller S, Mastrobuoni G, Hazapis O, Del GS, Schwefel D, Bley N, Voigt F, Chao J, Kempa S, H\u00fcttelmaier S, Selbach M, Landthaler M (2022). \u201cHDLBP binds ER-targeted mRNAs by multivalent interactions to promote protein synthesis of transmembrane and secreted proteins.\u201d Nat Commun, 13(1), 2727. doi:10.1038/s41467-022-30322-7 \u261e Link
D\u00f6rr D, Obermayer B, Weiner J, Zimmermann K, Anania C, Wagner L, Lyras E, Sapozhnikova V, Lara-Astiaso D, Pr\u00f3sper F, Lang R, Lupi\u00e1\u00f1ez D, Beule D, H\u00f6pken U, Leutz A, Mildner A (2022). \u201cC/EBP\u03b2 regulates lipid metabolism and Pparg isoform 2 expression in alveolar macrophages.\u201d Sci Immunol, 7(75), eabj0140. doi:10.1126/sciimmunol.abj0140 \u261e Link
Kedziora S, Obermayer B, Sugulle M, Herse F, Kr\u00e4ker K, Haase N, Langmia I, M\u00fcller D, Staff A, Beule D, Dechend R (2022). \u201cPlacental Transcriptome Profiling in Subtypes of Diabetic Pregnancies Is Strongly Confounded by Fetal Sex.\u201d Int J Mol Sci, 23(23). ISSN 1422-0067, doi:10.3390/ijms232315388 \u261e Link
Pennitz P, Kirsten H, Wyler E, Goekeri C, Obermayer B, Heinz GA, Mashreghi MF, Trimpert J, Landthaler M, Suttorp N, Hocke AC, Hippenstiel S, T\u00f6nnies M, Scholz M, Kuebler WM, Witzenrath M, Hoenzke K, Nouailles G (2022). \u201cIntegrational approaches for cross-species analysis of lung pathologies at single-cell resolution.\u201d Pneumologie, 76(S 01), FV 285. FV 285.
Hoffmann K, Obermayer B, H\u00f6nzke K, Fatykhova D, Demir Z, L\u00f6wa A, Alves L, Wyler E, Lopez-Rodriguez E, Mieth M, Baumgardt M, Hoppe J, Firsching T, T\u00f6nnies M, Bauer T, Eggeling S, Tran H, Schneider P, Neudecker J, R\u00fcckert J, Gruber A, Ochs M, Landthaler M, Beule D, Suttorp N, Hippenstiel S, Hocke A, Kessler M (2022). \u201cHuman alveolar progenitors generate dual lineage bronchioalveolar organoids.\u201d Commun Biol, 5(1), 875. ISSN 2399-3642, doi:10.1038/s42003-022-03828-5 \u261e Link
Neitzel H, Varon R, Chughtai S, Dartsch J, Dutrannoy-T\u00f6nsing V, N\u00fcrnberg P, N\u00fcrnberg G, Schweiger M, Digweed M, Hildebrand G, Hackmann K, Holtgrewe M, Sarioglu N, Schulze B, Horn D, Sperling K (2022). \u201cTransmission ratio distortion of mutations in the master regulator of centriole biogenesis PLK4.\u201d Hum Genet, 141(11), 1785-1794. doi:10.1007/s00439-022-02461-w \u261e Link
Foddis M, Blumenau S, Holtgrewe M, Paquette K, Westra K, Alonso I, Macario M, Morgadinho A, Velon A, Santo G, Santana I, M\u00f6nk\u00e4re S, Kuuluvainen L, Schleutker J, P\u00f6yh\u00f6nen M, Myllykangas L, Pavlovic A, Kostic V, Dobricic V, Lohmann E, Hanagasi H, Santos M, Guven G, Bilgic B, Bras J, Beule D, Dirnagl U, Guerreiro R, Sassi C (2022). \u201cTREX1 p.A129fs and p.Y305C variants in a large multi-ethnic cohort of CADASIL-like unrelated patients.\u201d doi:10.1016/j.neurobiolaging.2022.11.013 \u261e Link
Boschann F, Moreno D, Mensah M, Sczakiel H, Skipalova K, Holtgrewe M, Mundlos S, Fischer-Zirnsak B (2022). \u201cXq27.1 palindrome mediated interchromosomal insertion likely causes familial congenital bilateral laryngeal abductor paralysis (Plott syndrome).\u201d J Hum Genet, 67(7), 405-410. doi:10.1038/s10038-022-01018-z \u261e Link
Seidel F, Laser K, Klingel K, Dartsch J, Theisen S, Pickardt T, Holtgrewe M, G\u00e4rtner A, Berger F, Beule D, Milting H, Schubert S, Klaassen S, K\u00fchnisch J (2022). \u201cPathogenic Variants in Cardiomyopathy Disorder Genes Underlie Pediatric Myocarditis-Further Impact of Heterozygous Immune Disorder Gene Variants?\u201d J Cardiovasc Dev Dis, 9(7), 216. doi:10.3390/jcdd9070216 \u261e Link
May V, Koch L, Fischer-Zirnsak B, Horn D, Gehle P, Kornak U, Beule D, Holtgrewe M (2022). \u201cClearCNV: CNV calling from NGS panel data in the presence of ambiguity and noise.\u201d Bioinformatics, 38(16), 3871-3876. doi:10.1093/bioinformatics/btac418 \u261e Link
Sch\u00f6pflin R, Melo U, Moeinzadeh H, Heller D, Laupert V, Hertzberg J, Holtgrewe M, Alavi N, Klever M, Jungnitsch J, Comak E, T\u00fcrkmen S, Horn D, Duffourd Y, Faivre L, Callier P, Sanlaville D, Zuffardi O, Tenconi R, Kurtas N, Giglio S, Prager B, Latos-Bielenska A, Vogel I, Bugge M, Tommerup N, Spielmann M, Vitobello A, Kalscheuer V, Vingron M, Mundlos S (2022). \u201cIntegration of Hi-C with short and long-read genome sequencing reveals the structure of germline rearranged genomes.\u201d Nat Commun, 13(1), 6470. doi:10.1038/s41467-022-34053-7 \u261e Link
Rieke D, de BT, Horak P, Lamping M, Benary M, Jelas I, R\u00fcter G, Berger J, Zettwitz M, Kagelmann N, Kind A, Fabian F, Beule D, Glimm H, Brors B, Stenzinger A, Fr\u00f6hling S, Keilholz U (2022). \u201cFeasibility and outcome of reproducible clinical interpretation of high-dimensional molecular data: a comparison of two molecular tumor boards.\u201d BMC Med, 20, 367. ISSN 17417015, doi:10.1186/s12916-022-02560-5 \u261e Link
Hinze C, Kocks C, Leiz J, Karaiskos N, Boltengagen A, Cao S, Skopnik C, Klocke J, Hardenberg J, Stockmann H, Gotthardt I, Obermayer B, Haghverdi L, Wyler E, Landthaler M, Bachmann S, Hocke A, Corman V, Busch J, Schneider W, Himmerkus N, Bleich M, Eckardt K, Enghard P, Rajewsky N, Schmidt-Ott K (2022). \u201cSingle-cell transcriptomics reveals common epithelial response patterns in human acute kidney injury.\u201d Genome Med, 14, 103. doi:10.1186/s13073-022-01108-9 \u261e Link
Richter G, Wagner G, Reichenmiller K, Staufenbiel I, Martins O, L\u00f6scher B, Holtgrewe M, Jepsen S, Dommisch H, Schaefer A (2022). \u201cExome Sequencing of 5 Families with Severe Early-Onset Periodontitis.\u201d J Dent Res, 101(2), 151-157. doi:10.1177/00220345211029266 \u261e Link
Coutelier M, Holtgrewe M, J\u00e4ger M, Fl\u00f6ttman R, Mensah M, Spielmann M, Krawitz P, Horn D, Beule D, Mundlos S (2022). \u201cCombining callers improves the detection of copy number variants from whole-genome sequencing.\u201d Eur J Hum Genet, 30(2), 178-186. doi:10.1038/s41431-021-00983-x \u261e Link
Vogt G, Verheyen S, Schwartzmann S, Ehmke N, Potratz C, Schwerin-Nagel A, Plecko B, Holtgrewe M, Seelow D, Blatterer J, Speicher M, Kornak U, Horn D, Mundlos S, Fischer-Zirnsak B, Boschann F (2022). \u201cBiallelic truncating variants in ATP9A cause a novel neurodevelopmental disorder involving postnatal microcephaly and failure to thrive.\u201d J Med Genet, 59(7), 662-668. doi:10.1136/jmedgenet-2021-107843 \u261e Link
Kossen T, Hirzel MA, Madai VI, Boenisch F, Hennemuth A, Hildebrand K, Pokutta S, Sharma K, Hilbert A, Sobesky J, Galinovic I, Khalil AA, Fiebach JB and Frey D. Toward Sharing Brain Images: Differentially Private TOF-MRA Images With Segmentation Labels Using Generative Adversarial Networks. Frontiers in Artificial Intelligence. 5 (2022). issn: 2624-8212. doi: 10.3389/frai.2022.813842
"},{"location":"misc/publication-list/#2021","title":"2021","text":"Schinke C, Fernandez VV, Ivanov A, Peng Y, K\u00f6rtvelyessy P, Nolte L, Huehnchen P, Beule D, Stachelscheid H, Boehmerle W, Endres M (2021). \u201cDataset for: Modeling chemotherapy induced neurotoxicity with human induced pluripotent stem cell (iPSC)-derived sensory neurons.\u201d Data Brief, 38, 107320. doi:10.1016/j.dib.2021.107320 \u261e Link
Weiner J, Suwalski P, Holtgrewe M, Rakitko A, Thibeault C, M\u00fcller M, Patriki D, Quedenau C, Kr\u00fcger U, Ilinsky V, Popov I, Balnis J, Jaitovich A, Helbig E, Lippert L, Stubbemann P, Real L, Mac\u00edas J, Pineda J, Fernandez-Fuertes M, Wang X, Karadeniz Z, Saccomanno J, Doehn J, H\u00fcbner R, Hinzmann B, Salvo M, Blueher A, Siemann S, Jurisic S, Beer J, Rutishauser J, Wiggli B, Schmid H, Danninger K, Binder R, Corman V, M\u00fchlemann B, Arjun AR, Fragiadakis G, Mick E, Comet C, Calfee C, Erle D, Hendrickson C, Kangelaris K, Krummel M, Woodruff P, Langelier C, Venkataramani U, Garc\u00eda F, Zyla J, Drosten C, Alice B, Jones T, Suttorp N, Witzenrath M, Hippenstiel S, Zemojtel T, Skurk C, Poller W, Borodina T, Pa-Covid S, Ripke S, Sander L, Beule D, Landmesser U, Guettouche T, Kurth F, Heidecker B (2021). \u201cIncreased risk of severe clinical course of COVID-19 in carriers of HLA-C\\04:01.\u201d EClinicalMedicine, 40, 101099. doi:10.1016/j.eclinm.2021.101099 \u261e Link
Domaszewska T, Zyla J, Otto R, Kaufmann S, Weiner J (2021). \u201cGene Set Enrichment Analysis Reveals Individual Variability in Host Responses in Tuberculosis Patients.\u201d Front Immunol, 12, 694680. doi:10.3389/fimmu.2021.694680 \u261e Link
Nouailles G, Wyler E, Pennitz P, Postmus D, Vladimirova D, Kazmierski J, Pott F, Dietert K, Muelleder M, Farztdinov V, Obermayer B, Wienhold S, Andreotti S, Hoefler T, Sawitzki B, Drosten C, Sander LE, Suttorp N, Ralser M, Beule D, Gruber AD, Goffinet C, Landthaler M, Trimpert J, Witzenrath M (2021). \u201cSingle-cell-sequencing in SARS-COV-2-infected hamsters sheds light on endothelial cell involvement in COVID-19.\u201d European Respiratory Journal. doi:10.1183/13993003.congress-2021.PA2355 \u261e Link
Bischoff P, Trinks A, Obermayer B, Pett J, Wiederspahn J, Uhlitz F, Liang X, Lehmann A, Jurmeister P, Elsner A, Dziodzio T, R\u00fcckert J, Neudecker J, Falk C, Beule D, Sers C, Morkel M, Horst D, Bl\u00fcthgen N, Klauschen F (2021). \u201cSingle-cell RNA sequencing reveals distinct tumor microenvironmental patterns in lung adenocarcinoma.\u201d Oncogene, 40, 6748-6758. doi:10.1038/s41388-021-02054-3 \u261e Link
Uhlitz F, Bischoff P, Peidli S, Sieber A, Trinks A, L\u00fcthen M, Obermayer B, Blanc E, Ruchiy Y, Sell T, Mamlouk S, Arsie R, Wei T, Klotz-Noack K, Schwarz R, Sawitzki B, Kamphues C, Beule D, Landthaler M, Sers C, Horst D, Bl\u00fcthgen N, Morkel M (2021). \u201cMitogen-activated protein kinase activity drives cell trajectories in colorectal cancer.\u201d EMBO Mol Med, 13, e14123. doi:10.15252/emmm.202114123 \u261e Link
Nouailles G, Wyler E, Pennitz P, Postmus D, Vladimirova D, Kazmierski J, Pott F, Dietert K, Muelleder M, Farztdinov V, Obermayer B, Wienhold S, Andreotti S, Hoefler T, Sawitzki B, Drosten C, Sander L, Suttorp N, Ralser M, Beule D, Gruber A, Goffinet C, Landthaler M, Trimpert J, Witzenrath M (2021). \u201cTemporal omics analysis in Syrian hamsters unravel cellular effector responses to moderate COVID-19.\u201d Nat Commun, 12, 4869. doi:10.1038/s41467-021-25030-7 \u261e Link
Paterson S, Kar S, Ung S, Gardener Z, Bergstrom E, Ascough S, Kalyan M, Zyla J, Maertzdorf J, Mollenkopf H, Weiner J, Jozwik A, Jarvis H, Jha A, Nicholson B, Veldman T, Woods C, Mallia P, Kon O, Kaufmann S, Openshaw P, Chiu C (2021). \u201cInnate-like Gene Expression of Lung-Resident Memory CD8(+) T Cells during Experimental Human Influenza: A Clinical Study.\u201d Am J Respir Crit Care Med, 204, 826-841. doi:10.1164/rccm.202103-0620OC \u261e Link
Rajasekaran S, Siddiqui J, Rakijas J, Nicolay B, Lin C, Khan E, Patel R, Morris R, Wyler E, Boukhali M, Balasubramanyam J, Ranjith KR, van Rechem C, Vogel C, Elchuri S, Landthaler M, Obermayer B, Haas W, Dyson N, Miles W (2021). \u201cIntegrated multi-omics analysis of RB-loss identifies widespread cellular programming and synthetic weaknesses.\u201d Commun Biol, 4, 977. doi:10.1038/s42003-021-02495-2 \u261e Link
Schinke C, Fernandez VV, Ivanov A, Peng Y, K\u00f6rtvelyessy P, Nolte L, Huehnchen P, Beule D, Stachelscheid H, Boehmerle W, Endres M (2021). \u201cModeling chemotherapy induced neurotoxicity with human induced pluripotent stem cell (iPSC) -derived sensory neurons.\u201d Neurobiol Dis, 155, 105391. doi:10.1016/j.nbd.2021.105391 \u261e Link
Pei G, Zyla J, He L, Moura-Alves P, Steinle H, Saikali P, Lozza L, Nieuwenhuizen N, Weiner J, Mollenkopf H, Ellwanger K, Arnold C, Duan M, Dagil Y, Pashenkov M, Boneca I, Kufer T, Dorhoi A, Kaufmann S (2021). \u201cCellular stress promotes NOD\u00bd-dependent inflammation via the endogenous metabolite sphingosine-1-phosphate.\u201d EMBO J, 40, e106272. doi:10.15252/embj.2020106272 \u261e Link
Herda S, Heimann A, Obermayer B, Ciraolo E, Althoff S, Ru\u00df J, Grunert C, Busse A, Bullinger L, Pezzutto A, Blankenstein T, Beule D, Na I (2021). \u201cLong-term in vitro expansion ensures increased yield of central memory T cells as perspective for manufacturing challenges.\u201d Int J Cancer, 148, 3097-3110. doi:10.1002/ijc.33523 \u261e Link
Kraushar M, Krupp F, Harnett D, Turko P, Ambrozkiewicz M, Sprink T, Imami K, G\u00fcnnigmann M, Zinnall U, Vieira-Vieira C, Schaub T, M\u00fcnster-Wandowski A, B\u00fcrger J, Borisova E, Yamamoto H, Rasin M, Ohler U, Beule D, Mielke T, Tarabykin V, Landthaler M, Kramer G, Vida I, Selbach M, Spahn C (2021). \u201cProtein Synthesis in the Developing Neocortex at Near-Atomic Resolution Reveals Ebp1-Mediated Neuronal Proteostasis at the 60S Tunnel Exit.\u201d Mol Cell, 81, 304-322.e16. doi:10.1016/j.molcel.2020.11.037 \u261e Link
Heyckendorf J, Marwitz S, Reimann M, Avsar K, DiNardo A, G\u00fcnther G, Hoelscher M, Ibraim E, Kalsdorf B, Kaufmann S, Kontsevaya I, van Leth F, Mandalakas A, Maurer F, M\u00fcller M, Nitschkowski D, Olaru I, Popa C, Rachow A, Rolling T, Rybniker J, Salzer H, Sanchez-Carballo P, Schuhmann M, Schaub D, Spinu V, Su\u00e1rez I, Terhalle E, Unnewehr M, Weiner J3, Goldmann T, Lange C (2021). \u201cPrediction of anti-tuberculosis treatment duration based on a 22-gene transcriptomic model.\u201d Eur Respir J, 58. doi:10.1183/13993003.03492-2020 \u261e Link
Bao X, Weiner J3, Meckes O, Dommisch H, Schaefer A (2021). \u201cEntamoeba gingivalis Exerts Severe Pathogenic Effects on the Oral Mucosa.\u201d J Dent Res, 100, 771-776. doi:10.1177/00220345211004498 \u261e Link
Eckold C, Kumar V, Weiner J, Alisjahbana B, Riza A, Ronacher K, Coronel J, Kerry-Barnard S, Malherbe S, Kleynhans L, Stanley K, Ruslami R, Ioana M, Ugarte-Gil C, Walzl G, van Crevel R, Wijmenga C, Critchley J, Dockrell H, Cliff J, TANDEM c (2021). \u201cImpact of Intermediate Hyperglycemia and Diabetes on Immune Dysfunction in Tuberculosis.\u201d Clin Infect Dis, 72, 69-78. doi:10.1093/cid/ciaa751 \u261e Link
Wyler E, M\u00f6sbauer K, Franke V, Diag A, Gottula L, Arsi\u00e8 R, Klironomos F, Koppstein D, H\u00f6nzke K, Ayoub S, Buccitelli C, Hoffmann K, Richter A, Legnini I, Ivanov A, Mari T, Del GS, Papies J, Praktiknjo S, Meyer T, M\u00fcller M, Niemeyer D, Hocke A, Selbach M, Akalin A, Rajewsky N, Drosten C, Landthaler M (2021). \u201cTranscriptomic profiling of SARS-CoV-2 infected human cell lines identifies HSP90 as target for COVID-19 therapy.\u201d iScience, 24, 102151. doi:10.1016/j.isci.2021.102151 \u261e Link
M\u00f6lder F, Jablonski K, Letcher B, Hall M, Tomkins-Tinch C, Sochat V, Forster J, Lee S, Twardziok S, Kanitz A, Wilm A, Holtgrewe M, Rahmann S, Nahnsen S, K\u00f6ster J (2021). \u201cSustainable data analysis with Snakemake.\u201d F1000Res, 10, 33. doi:10.12688/f1000research.29032.2 \u261e Link
Messerschmidt C, Foddis M, Blumenau S, M\u00fcller S, Bentele K, Holtgrewe M, Kun-Rodrigues C, Alonso I, do CMM, Morgadinho A, Velon A, Santo G, Santana I, M\u00f6nk\u00e4re S, Kuuluvainen L, Schleutker J, P\u00f6yh\u00f6nen M, Myllykangas L, Senatore A, Berchtold D, Winek K, Meisel A, Pavlovic A, Kostic V, Dobricic V, Lohmann E, Hanagasi H, Guven G, Bilgic B, Bras J, Guerreiro R, Beule D, Dirnagl U, Sassi C (2021). \u201cPHACTR1 genetic variability is not critical in small vessel ischemic disease patients and PcomA recruitment in C57BL/6J mice.\u201d Sci Rep, 11(1), 6072. doi:10.1038/s41598-021-84919-x \u261e Link
Zarate Y, Uehara T, Abe K, Oginuma M, Harako S, Ishitani S, Lehesjoki A, Bierhals T, Kloth K, Ehmke N, Horn D, Holtgrewe M, Anderson K, Viskochil D, Edgar-Zarate C, Sacoto M, Schnur R, Morrow M, Sanchez-Valle A, Pappas J, Rabin R, Muona M, Anttonen A, Platzer K, Luppe J, Gburek-Augustat J, Kaname T, Okamoto N, Mizuno S, Kaido Y, Ohkuma Y, Hirose Y, Ishitani T, Kosaki K (2021). \u201cCDK19-related disorder results from both loss-of-function and gain-of-function de novo missense variants.\u201d Genet Med, 23(6), 1050-1057. doi:10.1038/s41436-020-01091-9 \u261e Link
Vogt G, El CN, Herczegfalvi \u00c1, K\u00f6lbel H, Lekaj A, Schara U, Holtgrewe M, Krause S, Horvath R, Schuelke M, H\u00fcbner C, Mundlos S, Roos A, Lochm\u00fcller H, Karcagi V, Kornak U, Fischer-Zirnsak B (2021). \u201cExpanding the clinical and molecular spectrum of ATP6V1A related metabolic cutis laxa.\u201d J Inherit Metab Dis, 44(4), 972-986. doi:10.1002/jimd.12341 \u261e Link
Yildirim F, Foddis M, Blumenau S, M\u00fcller S, Kajetan B, Holtgrewe M, Kola V, Beule D, Sassi C (2021). \u201cShared and oppositely regulated transcriptomic signatures in Huntington's disease and brain ischemia confirm known and unveil novel potential neuroprotective genes.\u201d Neurobiol Aging, 104, 122.e1-122.e17. doi:10.1016/j.neurobiolaging.2021.03.001 \u261e Link
Elsner J, Mensah M, Holtgrewe M, Hertzberg J, Bigoni S, Busche A, Coutelier M, de SD, El\u00e7ioglu N, Filges I, Gerkes E, Girisha K, Graul-Neumann L, Jamsheer A, Krawitz P, Kurth I, Markus S, Megarbane A, Reis A, Reuter M, Svoboda D, Teller C, Tuysuz B, T\u00fcrkmen S, Wilson M, Woitschach R, Vater I, Caliebe A, H\u00fclsemann W, Horn D, Mundlos S, Spielmann M (2021). \u201cGenome sequencing in families with congenital limb malformations.\u201d Hum Genet, 140(8), 1229-1239. doi:10.1007/s00439-021-02295-y \u261e Link
Seidel F, Holtgrewe M, Al-Wakeel-Marquard N, Opgen-Rhein B, Dartsch J, Herbst C, Beule D, Pickardt T, Klingel K, Messroghli D, Berger F, Schubert S, K\u00fchnisch J, Klaassen S (2021). \u201cPathogenic Variants Associated With Dilated Cardiomyopathy Predict Outcome in Pediatric Myocarditis.\u201d Circ Genom Precis Med, 14(4), e003250. doi:10.1161/CIRCGEN.120.003250 \u261e Link
Melo U, Piard J, Fischer-Zirnsak B, Klever M, Sch\u00f6pflin R, Mensah M, Holtgrewe M, Arbez-Gindre F, Martin A, Guigue V, Gaillard D, Landais E, Roze V, Kremer V, Ramanah R, Cabrol C, Harms F, Kornak U, Spielmann M, Mundlos S, van Maldergem L (2021). \u201cComplete lung agenesis caused by complex genomic rearrangements with neo-TAD formation at the SHH locus.\u201d Hum Genet, 140(10), 1459-1469. doi:10.1007/s00439-021-02344-6 \u261e Link
Sczakiel H, H\u00fclsemann W, Holtgrewe M, Abad-Perez A, Elsner J, Schwartzmann S, Horn D, Spielmann M, Mundlos S, Mensah M (2021). \u201cGLI3 variants causing isolated polysyndactyly are not restricted to the protein's C-terminal third.\u201d Clin Genet, 100(6), 758-765. doi:10.1111/cge.14059 \u261e Link
Li, N., Hollunder, B., Baldermann, J. C., Kibleur, A., Treu, S., Akram, H., Al-Fatly, B., Strange, B. A., Barcia, J. A., Zrinzo, L., Joyce, E. M., Chabardes, S., Visser-Vandewalle, V., Polosan, M., Kuhn, J., K\u00fchn, A. A., & Horn, A. (2021). A Unified Functional Network Target for Deep Brain Stimulation in Obsessive-Compulsive Disorder. Biological Psychiatry. doi: 10.1016/j.biopsych.2021.04.006
Bressem KK, Vahldiek JL, Adams L, Niehues SM, Haibel H, Rodriguez VR, Torgutalp M, Protopopov M, Proft F, Rademacher J, Sieper J, Rudwaleit M, Hamm B, Makowski MR, Hermann KG, Poddubnyy D. Deep learning for detection of radiographic sacroiliitis: achieving expert-level performance. Arthritis Res Ther. 2021 Apr 8;23(1):106. doi: 10.1186/s13075-021-02484-0
Kossen T, Subramaniam P, Madai VI, Hennemuth A, Hildebrand K, Hilbert A, Sobesky J, Livne M, Galinovic I, Khalil AA, Fiebach JB, Frey D. Synthesizing anonymized and labeled TOF-MRA patches for brain vessel segmentation using generative adversarial networks. Computers in Biology and Medicine. 2021 Apr 131,104254. doi: 10.1016/j.compbiomed.2021.104254
Paraskevopoulou S., K\u00e4fer S., Zirkel F., Donath A., Petersen M., Liu S., Zhou X., Drosten C., Misof B., Junglen S. (2021). \"Viromics of extant insect orders unveil the evolution of the flavi-like superfamily.\" Virus Evolution 2021 Mar 30. doi: 10.1093/ve/veab030
Thomas Krannich, W Timothy J White, Sebastian Niehus, Guillaume Holley, Bjarni V Halld\u00f3rsson, Birte Kehr, Population-scale detection of non-reference sequence variants using colored de Bruijn graphs, Bioinformatics, 2021, btab749, doi: 10.1093/bioinformatics/btab749
Julia Markowski, Rieke Kempfer, Alexander Kukalev, Ibai Irastorza-Azcarate, Gesa Loof, Birte Kehr, Ana Pombo, Sven Rahmann, Roland F Schwarz, GAMIBHEAR: whole-genome haplotype reconstruction from Genome Architecture Mapping data, Bioinformatics, Volume 37, Issue 19, 1 October 2021, Pages 3128\u20133135. doi: 10.1093/bioinformatics/btab238
"},{"location":"misc/publication-list/#2020","title":"2020","text":"Kr\u00fctzfeldt LM, Schubach M, Kircher M. The impact of different negative training data on regulatory sequence predictions. PLoS One. 2020 Dec 1;15(12):e0237412. doi: 10.1371/journal.pone.0237412.
Klotz-Noack K, Klinger B, Rivera M, Bublitz N, Uhlitz F, Riemer P, L\u00fcthen M, Sell T, Kasack K, Gastl B, Ispasanie SSS, Simon T, Janssen N, Schwab M, Zuber J, Horst D, Bl\u00fcthgen N, Sch\u00e4fer R, Morkel M, Sers C. SFPQ Depletion Is Synthetically Lethal with BRAFV600E in Colorectal Cancer Cells. Cell Rep. 2020 Sep 22;32(12):108184. doi: 10.1016/j.celrep.2020.108184.
Kleinert, P., Martin, B., & Kircher, M. (2020). \"HemoMIPs\u2014Automated analysis and result reporting pipeline for targeted sequencing data.\" PLOS Computational Biology, 16(6), e1007956. doi: 10.1371/journal.pcbi.1007956
Ehmke, N.; Cusmano-Ozog, K.; Koenig, R.; Holtgrewe, M.; Nur, B.; Mihci, E.; Babcock, H.; Gonzaga-Jauregui, C.; Overton, J. D.; Xiao, J.; et al. Biallelic Variants in KYNU Cause a Multisystemic Syndrome with Hand Hyperphalangism. Bone 2020, 115219. doi: 10.1016/j.bone.2019.115219.
Niehus, S.; J\u00f3nsson, H.; Sch\u00f6nberger, J.; Bj\u00f6rnsson, E.; Beyter, D.; Eggertsson, H.P.; Sulem, P.; Stef\u00e1nsson, K.; Halld\u00f3rsson, B.V.; Kehr, B. PopDel identifies medium-size deletions jointly in tens of thousands of genomes. bioRxiv 2020, 10.1101/740225 doi: 10.1101/740225
Gordon, M. G., Inoue, F., Martin, B., Schubach, M., Agarwal, V., Whalen, S., ... & Kreimer, A. (2020). \"lentiMPRA and MPRAflow for high-throughput functional characterization of gene regulatory elements.\" Nature Protocols, 15(8), 2387-2412. doi: 10.1038/s41596-020-0333-5
Paraskevopoulou S., Pirzer F., Goldmann N., Schmid J., Corman V.M., Gottula L.T.,Schroeder S., Rasche A., Muth D., Drexler J.F., Heni A.C., Eibner G.J., Page R.A., Jones T.C., M\u00fcllerM.A., Sommer S., Glebe D., and Drosten C. (2020). \"Mammalian deltavirus without hepadnavirus coinfection in the neotropical rodent Proechimys semispinosus.\" Proceedings of the National Academy of Sciences 2020 Jul 28;117(30):17977-17983. doi: 10.1073/pnas.2006750117.
Mattei D, Ivanov A, van Oostrum M, Pantelyushin S, Richetto J, Mueller F, Beffinger M, Schellhammer L, Vom BJ, Wollscheid B, Beule D, Paolicelli R, Meyer U (2020). \u201cEnzymatic Dissociation Induces Transcriptional and Proteotype Bias in Brain Cell Populations.\u201d Int J Mol Sci, 21. doi:10.3390/ijms21217944 \u261e Link
Kuhring M, Eisenberger A, Schmidt V, Kr\u00e4nkel N, Leistner D, Kirwan J, Beule D (2020). \u201cConcepts and Software Package for Efficient Quality Control in Targeted Metabolomics Studies: MeTaQuaC.\u201d Anal Chem, 92, 10241-10245. doi:10.1021/acs.analchem.0c00136 \u261e Link
Koche R, Rodriguez-Fos E, Helmsauer K, Burkert M, MacArthur I, Maag J, Chamorro R, Munoz-Perez N, Puiggr\u00f2s M, Dorado GH, Bei Y, R\u00f6efzaad C, Bardinet V, Szymansky A, Winkler A, Thole T, Timme N, Kasack K, Fuchs S, Klironomos F, Thiessen N, Blanc E, Schmelz K, K\u00fcnkele A, Hundsd\u00f6rfer P, Rosswog C, Theissen J, Beule D, Deubzer H, Sauer S, Toedling J, Fischer M, Hertwig F, Schwarz R, Eggert A, Torrents D, Schulte J, Henssen A (2020). \u201cExtrachromosomal circular DNA drives oncogenic genome remodeling in neuroblastoma.\u201d Nat Genet, 52, 29-34. doi:10.1038/s41588-019-0547-z \u261e Link
Holtgrewe M, Stolpe O, Nieminen M, Mundlos S, Knaus A, Kornak U, Seelow D, Segebrecht L, Spielmann M, Fischer-Zirnsak B, Boschann F, Scholl U, Ehmke N, Beule D (2020). \u201cVarFish: comprehensive DNA variant analysis for diagnostics and research.\u201d Nucleic Acids Res, 48, W162-W169. doi:10.1093/nar/gkaa241 \u261e Link
Obermayer B, Holtgrewe M, Nieminen M, Messerschmidt C, Beule D (2020). \u201cSCelVis: exploratory single cell data analysis on the desktop and in the cloud.\u201d PeerJ, 8, e8607. doi:10.7717/peerj.8607 \u261e Link
Nieminen M, Stolpe O, Schumann F (2020). \u201cSODAR Core: a Django-based framework for scientific data management and analysis web apps.\u201d Journal of Open Source Software, 5(55), 1584. doi:10.21105/joss.01584 \u261e Link
Akpa C, Kleo K, Oker E, Tomaszewski N, Messerschmidt C, L\u00f3pez C, Wagener R, Oehl-Huber K, Dettmer K, Schoeler A, Lenze D, Oefner P, Beule D, Siebert R, Capper D, Dimitrova L, Hummel M (2020). \u201cAcquired resistance to DZNep-mediated apoptosis is associated with copy number gains of AHCY in a B-cell lymphoma model.\u201d BMC Cancer, 20, 427. doi:10.1186/s12885-020-06937-8 \u261e Link
Messerschmidt C, Obermayer B, Klinghammer K, Ochsenreither S, Treue D, Stenzinger A, Glimm H, Fr\u00f6hling S, Kindler T, Brandts C, Schulze-Osthoff K, Weichert W, Tinhofer I, Klauschen F, Keilholz U, Beule D, Rieke D (2020). \u201cDistinct immune evasion in APOBEC-enriched, HPV-negative HNSCC.\u201d Int J Cancer, 147, 2293-2302. doi:10.1002/ijc.33123 \u261e Link
Boschann F, Fischer-Zirnsak B, Wienker T, Holtgrewe M, Seelow D, Eichhorn B, D\u00f6hnert S, Fahsold R, Horn D, Graul-Neumann L (2020). \u201cAn intronic splice site alteration in combination with a large deletion affecting VPS13B (COH1) causes Cohen syndrome.\u201d Eur J Med Genet, 63, 103973. doi:10.1016/j.ejmg.2020.103973 \u261e Link
Rieke D, Klinghammer K, Obermayer B, Blanc E, Messerschmidt C, J\u00f6hrens K, Klauschen F, Tinhofer I, Beule D, Keilholz U, Ochsenreither S (2020). \u201cAssociation of a STK11/KEAP1-Mutation Gene Expression Signature in Lung Adenocarcinoma with Immune Desertion in Squamous Cell Carcinomas and Mediation by NFE2L2 Deregulation.\u201d J Clin Oncol, 38(15_suppl), 3082-3082. doi:10.1200/JCO.2020.38.15_suppl.3082 \u261e Link
Ehmke N, Cusmano-Ozog K, Koenig R, Holtgrewe M, Nur B, Mihci E, Babcock H, Gonzaga-Jauregui C, Overton J, Xiao J, Martinez A, Muenke M, Balzer A, Jochim J, El CN, Fischer-Zirnsak B, Huber C, Kornak U, Elsea S, Cormier-Daire V, Ferreira C (2020). \u201cBiallelic variants in KYNU cause a multisystemic syndrome with hand hyperphalangism.\u201d Bone, 133, 115219. doi:10.1016/j.bone.2019.115219 \u261e Link
Seidel E, Walenda G, Messerschmidt C, Obermayer B, Peitzsch M, Wallace P, Bahethi R, Yoo T, Choi M, Schrade P, Bachmann S, Liebisch G, Eisenhofer G, Beule D, Scholl U (2020). \u201cGeneration and characterization of a mitotane-resistant adrenocortical cell line.\u201d Endocr Connect, 9, 122-134. doi:10.1530/EC-19-0510 \u261e Link
Melo U, Sch\u00f6pflin R, Acuna-Hidalgo R, Mensah M, Fischer-Zirnsak B, Holtgrewe M, Klever M, T\u00fcrkmen S, Heinrich V, Pluym I, Matoso E, Bernardo dSS, Louro P, H\u00fclsemann W, Cohen M, Dufke A, Latos-Biele\u0144ska A, Vingron M, Kalscheuer V, Quintero-Rivera F, Spielmann M, Mundlos S (2020). \u201cHi-C Identifies Complex Genomic Rearrangements and TAD-Shuffling in Developmental Diseases.\u201d Am J Hum Genet, 106, 872-884. doi:10.1016/j.ajhg.2020.04.016 \u261e Link
Robinson P, Ravanmehr V, Jacobsen J, Danis D, Zhang X, Carmody L, Gargano M, Thaxton C, UNC BC, Karlebach G, Reese J, Holtgrewe M, K\u00f6hler S, McMurry J, Haendel M, Smedley D (2020). \u201cInterpretable Clinical Genomics with a Likelihood Ratio Paradigm.\u201d Am J Hum Genet, 107, 403-417. doi:10.1016/j.ajhg.2020.06.021 \u261e Link
Blumenau S, Foddis M, M\u00fcller S, Holtgrewe M, Bentele K, Berchtold D, Beule D, Dirnagl U, Sassi C (2020). \u201cInvestigating APOE, APP-A\u03b2 metabolism genes and Alzheimer's disease GWAS hits in brain small vessel ischemic disease.\u201d Sci Rep, 10, 7103. doi:10.1038/s41598-020-63183-5 \u261e Link
Quade A, Thiel A, Kurth I, Holtgrewe M, Elbracht M, Beule D, Eggermann K, Scholl U, H\u00e4usler M (2020). \u201cParoxysmal tonic upgaze: A heterogeneous clinical condition responsive to carbonic anhydrase inhibition.\u201d Eur J Paediatr Neurol, 25, 181-186. doi:10.1016/j.ejpn.2019.11.002 \u261e Link
Thole T, Toedling J, Spr\u00fcssel A, Pfeil S, Savelyeva L, Capper D, Messerschmidt C, Beule D, Groeneveld-Krentz S, Eckert C, Gambara G, Henssen A, Finkler S, Schulte J, Sieber A, Bluethgen N, Regenbrecht C, K\u00fcnkele A, Lodrini M, Eggert A, Deubzer H (2020). \u201cReflection of neuroblastoma intratumor heterogeneity in the new OHC-NB1 disease model.\u201d Int J Cancer, 146, 1031-1041. doi:10.1002/ijc.32572 \u261e Link
Lamping M, Benary M, Leyvraz S, Messerschmidt C, Blanc E, Kessler T, Sch\u00fctte M, Lenze D, J\u00f6hrens K, Burock S, Klinghammer K, Ochsenreither S, Sers C, Sch\u00e4fer R, Tinhofer I, Beule D, Klauschen F, Yaspo M, Keilholz U, Rieke D (2020). \u201cSupport of a molecular tumour board by an evidence-based decision management system for precision oncology.\u201d Eur J Cancer, 127, 41-51. doi:10.1016/j.ejca.2019.12.017 \u261e Link
Praktiknjo S, Obermayer B, Zhu Q, Fang L, Liu H, Quinn H, Stoeckius M, Kocks C, Birchmeier W, Rajewsky N (2020). \u201cTracing tumorigenesis in a solid tumor model at single-cell resolution.\u201d Nat Commun, 11, 991. doi:10.1038/s41467-020-14777-0 \u261e Link
Suter A, Santos-Simarro F, Toerring P, Abad PA, Ramos-Mejia R, Heath K, Huckstadt V, Parr\u00f3n-Pajares M, Mensah M, H\u00fclsemann W, Holtgrewe M, Mundlos S, Kornak U, Bartsch O, Ehmke N (2020). \u201cVariable pulmonary manifestations in Chitayat syndrome: Six additional affected individuals.\u201d Am J Med Genet A, 182, 2068-2076. doi:10.1002/ajmg.a.61735 \u261e Link
Bockstahler M, Fischer A, Goetzke C, Neumaier H, Sauter M, Kespohl M, M\u00fcller A, Meckes C, Salbach C, Schenk M, Heuser A, Landmesser U, Weiner J, Meder B, Lehmann L, Kratzer A, Klingel K, Katus H, Kaya Z, Beling A (2020). \u201cHeart-Specific Immune Responses in an Animal Model of Autoimmune-Related Myocarditis Mitigated by an Immunoproteasome Inhibitor and Genetic Ablation.\u201d Circulation, 141, 1885-1902. doi:10.1161/CIRCULATIONAHA.119.043171 \u261e Link
Krishnamoorthy G, Kaiser P, Abu AU, Weiner J3, Moura-Alves P, Brinkmann V, Kaufmann S (2020). \u201cFX11 limits Mycobacterium tuberculosis growth and potentiates bactericidal activity of isoniazid through host-directed activity.\u201d Dis Model Mech, 13. doi:10.1242/dmm.041954 \u261e Link
Tregoning J, Weiner J, Cizmeci D, Hake D, Maertzdorf J, Kaufmann S, Leroux-Roels G, Maes C, Aerssens A, Calvert A, Jones C (2020). \u201cPregnancy has a minimal impact on the acute transcriptional signature to vaccination.\u201d NPJ Vaccines, 5, 29. doi:10.1038/s41541-020-0177-6 \u261e Link
Weiner J, Domaszewska T, Donkor S, Kaufmann S, Hill P, Sutherland J (2020). \u201cChanges in Transcript, Metabolite, and Antibody Reactivity During the Early Protective Immune Response in Humans to Mycobacterium tuberculosis Infection.\u201d Clin Infect Dis, 71, 30-40. doi:10.1093/cid/ciz785 \u261e Link
Win Z, Weiner J, Listanco A, Patel N, Sharma R, Greenwood A, Maertzdorf J, Mollenkopf H, Pizzoferro K, Cole T, Bodinham C, Kaufmann S, Denoel P, Del GG, Lewis D (2020). \u201cSystematic Evaluation of Kinetics and Distribution of Muscle and Lymph Node Activation Measured by (18)F-FDG- and (11)C-PBR28-PET/CT Imaging, and Whole Blood and Muscle Transcriptomics After Immunization of Healthy Humans With Adjuvanted and Unadjuvanted Vaccines.\u201d Front Immunol, 11, 613496. doi:10.3389/fimmu.2020.613496 \u261e Link
Mochel F, Rastetter A, Ceulemans B, Platzer K, Yang S, Shinde D, Helbig K, Lopergolo D, Mari F, Renieri A, Benetti E, Canitano R, Waisfisz Q, Plomp A, Huisman S, Wilson G, Cathey S, Louie R, Gaudio D, Waggoner D, Kacker S, Nugent K, Roeder E, Bruel A, Thevenon J, Ehmke N, Horn D, Holtgrewe M, Kaiser F, Kamphausen S, Abou JR, Weckhuysen S, Dalle C, Depienne C (2020). \u201cVariants in the SK2 channel gene (KCNN2) lead to dominant neurodevelopmental movement disorders.\u201d Brain, 143(12), 3564-3573. doi:10.1093/brain/awaa346 \u261e Link
Foddis M, Winek K, Bentele K, Mueller S, Blumenau S, Reichhart NN, Crespo-Garcia S, Harnett D, Ivanov A, Meisel A, Joussen A, Strauss O, Beule D, Dirnagl U, Sassi C (2020). \u201cAn exploratory investigation of brain collateral circulation plasticity after cerebral ischemia in two experimental C57BL/6 mouse models.\u201d J Cereb Blood Flow Metab, 40, 276-287. doi:10.1177/0271678X19827251 \u261e Link
Quade A, Thiel A, Kurth I, Holtgrewe M, Elbracht M, Beule D, Eggermann K, Scholl U, H\u00e4usler M (2020). \u201cParoxysmal tonic upgaze: A heterogeneous clinical condition responsive to carbonic anhydrase inhibition.\u201d Eur J Paediatr Neurol, 25, 181-186. doi:10.1016/j.ejpn.2019.11.002 \u261e Link
"},{"location":"misc/publication-list/#2019","title":"2019","text":"Kircher, M., Xiong, C., Martin, B. et al. \"Saturation mutagenesis of twenty disease-associated regulatory elements at single base-pair resolution.\" Nat Commun 10, 3583 (2019). doi: 10.1038/s41467-019-11526-w
Stefanovski L, Triebkorn P, Spiegler A, Diaz-Cortes M-A, Solodkin A, Jirsa V, McIntosh RA and Ritter P (2019). \"Linking Molecular Pathways and Large-Scale Computational Modeling to Assess Candidate Disease Mechanisms and Pharmacodynamics in Alzheimer's Disease.\" Front. Comput. Neurosci.. 13:54. doi: 10.3389/fncom.2019.00054
Boeddrich A., Babila J.T., Wiglenda T., Diez L., Jacob M., Nietfeld W., Huska M.R., Haenig C., Groenke N., Buntru A., Blanc E., Meier J.C., Vannoni E., Erck C., Friedrich B., Martens H., Neuendorf N., Schnoegl S., Wolfer DP., Loos M., Beule D., Andrade-Navarro M.A., Wanker E.E. (2019). \"The Anti-amyloid Compound DO1 Decreases Plaque Pathology and Neuroinflammation-Related Expression Changes in 5xFAD Transgenic Mice.\" Cell Chem Biol. 2019 Jan 17;26(1):109-120.e7. doi: 10.1016/j.chembiol.2018.10.013.
Fountain M.D., Oleson, D.S., Rech. M.E., Segebrecht, L., Hunter, J.V., McCarthy, J.M., Lupo, P.J., Holtgrewe, M., Mora, R., Rosenfeld, J.A., Isidor, B., Le Caignec, C., Saenz, M.S., Pedersen, R.C., Morgen, T.M., Pfotenhauer, J.P., Xia, F., Bi, W., Kang, S.-H.L., Patel, A., Krantz, I.D., Raible, S.E., Smith, W.E., Cristian, I., Tori, E., Juusola, J., Millan, F., Wentzensen, I.M., Person, R.E., K\u00fcry, S., B\u00e9zieau, S., Uguen, K., F\u00e9rec, C., Munnich, A., van Haelst, M., Lichtenbelt, K.D., van Gassen, K., Hagelstrom, T., Chawla, A., Perry, D.L., Taft, R.J., Jones, M., Masser-Frye, D., Dyment, D., Venkateswaran, S., Li, C., Escobar, L,.F., Horn, D., Spillmann, R.C., Pe\u00f1a, L., Wierzba, J., Strom, T.M. Parent, I. Kaiser, F.J., Ehmke, N., Schaaf, C.P. (2019). \"Pathogenic variants in USP7 cause a neurodevelopmental disorder with speech delays, altered behavior, and neurologic anomalies.\" Genet. Med. 2019 Jan 25. doi: 10.1038/s41436-019-0433-1
Holtgrewe,M., Messerschmidt,C., Nieminen,M. and Beule,D. (2019) DigestiFlow: from BCL to FASTQ with ease. Bioinformatics, 10.1093/bioinformatics/btz850.
K\u00e4fer S., Paraskevopoulou S., Zirkel F., Wieseke N., Donath A., Petersen M., Jones T.C., Liu S., Zhou X., Middendorf M., Junglen S., Misof B., Drosten C. (2019). \"Re-assessing the diversity of negative strand RNA viruses in insects.\" PLOS Pathogens 2019 Dec 12. doi: 10.1371/journal.ppat.1008224
K\u00fchnisch,J., Herbst,C., Al\u2010Wakeel\u2010Marquard,N., Dartsch,J., Holtgrewe,M., Baban,A., Mearini,G., Hardt,J., Kolokotronis,K., Gerull,B., et al. (2019) Targeted panel sequencing in pediatric primary cardiomyopathy supports a critical role of TNNI3. Clin Genet, 96, 549\u2013559. https://doi.org/10.1111/cge.13645
Marklewitz M., Dutari L.C., Paraskevopoulou S., Page R.A., Loaiza J.R., Junglen S. (2019). \"Diverse novel phleboviruses in sandflies from the Panama Canal area, Central Panama.\" Journal of General Virology 2019 May 3. doi: 10.1099/jgv.0.001260
Quade,A., Thiel,A., Kurth,I., Holtgrewe,M., Elbracht,M., Beule,D., Eggermann,K., Scholl,U.I. and H\u00e4usler,M. (2019) Paroxysmal tonic upgaze: A heterogeneous clinical condition responsive to carbonic anhydrase inhibition. European Journal of Paediatric Neurology, 10.1016/j.ejpn.2019.11.002.
Holtgrewe M, Messerschmidt C, Nieminen M, Beule D (2019). \u201cDigestiflow: From BCL to FASTQ with Ease.\u201d Bioinformatics.
Schumann F, Blanc E, Messerschmidt C, Blankenstein T, Busse A, Beule D (2019). \u201cSigsPack, a package for cancer mutational signatures.\u201d BMC Bioinformatics, 20, 450. doi:10.1186/s12859-019-3043-7 \u261e Link
Blanc E, Holtgrewe M, Dhamodaran A, Messerschmidt C, Willimsky G, Blankenstein T, Beule D (2019). \u201cIdentification and ranking of recurrent neo-epitopes in cancer.\u201d BMC Med Genomics, 12, 171. doi:10.1186/s12920-019-0611-7 \u261e Link
Boeddrich A, Babila J, Wiglenda T, Diez L, Jacob M, Nietfeld W, Huska M, Haenig C, Groenke N, Buntru A, Blanc E, Meier J, Vannoni E, Erck C, Friedrich B, Martens H, Neuendorf N, Schnoegl S, Wolfer D, Loos M, Beule D, Andrade-Navarro M, Wanker E (2019). \u201cThe Anti-amyloid Compound DO1 Decreases Plaque Pathology and Neuroinflammation-Related Expression Changes in 5xFAD Transgenic Mice.\u201d Cell Chem Biol, 26, 109-120.e7. doi:10.1016/j.chembiol.2018.10.013 \u261e Link
Borgsm\u00fcller N, Gloaguen Y, Opialla T, Blanc E, Sicard E, Royer A, Le BB, Durand S, Mign\u00e9 C, P\u00e9t\u00e9ra M, Pujos-Guillot E, Giacomoni F, Guitton Y, Beule D, Kirwan J (2019). \u201cWiPP: Workflow for Improved Peak Picking for Gas Chromatography-Mass Spectrometry (GC-MS) Data.\u201d Metabolites, 9. doi:10.3390/metabo9090171 \u261e Link
Buonfiglioli A, Efe I, Guneykaya D, Ivanov A, Huang Y, Orlowski E, Kr\u00fcger C, Deisz R, Markovic D, Fl\u00fch C, Newman A, Schneider U, Beule D, Wolf S, Dzaye O, Gutmann D, Semtner M, Kettenmann H, Lehnardt S (2019). \u201clet-7 MicroRNAs Regulate Microglial Function and Suppress Glioma Growth through Toll-Like Receptor 7.\u201d Cell Rep, 29, 3460-3471.e7. doi:10.1016/j.celrep.2019.11.029 \u261e Link
K\u00fchnisch J, Herbst C, Al-Wakeel-Marquard N, Dartsch J, Holtgrewe M, Baban A, Mearini G, Hardt J, Kolokotronis K, Gerull B, Carrier L, Beule D, Schubert S, Messroghli D, Degener F, Berger F, Klaassen S (2019). \u201cTargeted panel sequencing in pediatric primary cardiomyopathy supports a critical role of TNNI3.\u201d Clin Genet, 96, 549-559. doi:10.1111/cge.13645 \u261e Link
Kuhring M, Nieminen M, Kirwan J, Beule D, Holtgrewe M (2019). \u201cAltamISA: A Python API for ISA-Tab Files.\u201d JOSS, 4, 1610. doi:10.21105/joss.01610
Lamping M, Rieke D, Klauschen F, J\u00f6hrens K, Anagnostopoulos I, Lenze D, Tinhofer I, Benary M, Ochsenreither S, Klinghammer K, Burock S, Jann H, St\u00fcven A, Ditzen D, Beule D, Messerschmidt C, Blanc E, Sch\u00e4fer R, Keilholz U (2019). \u201cClinical Impact of Comprehensive versus Targeted Genomic Analysis for Precision Oncology.\u201d J Clin Oncol, 37(15_suppl), e13033-e13033. doi:10.1200/JCO.2019.37.15_suppl.e13033 \u261e Link
Herda S, Heimann A, Althoff S, Ru\u00df J, Bullinger L, Beule D, Obermayer B, Na I (2019). \u201cLong-Term T Cell Expansion Results in Increased Numbers of Central Memory T Cells with Sustained Functional Properties for Adoptive T Cell Therapy.\u201d Blood. doi:10.1182/blood-2019-124736 \u261e Link
van Heesch S, Witte F, Schneider-Lunitz V, Schulz J, Adami E, Faber A, Kirchner M, Maatz H, Blachut S, Sandmann C, Kanda M, Worth C, Schafer S, Calviello L, Merriott R, Patone G, Hummel O, Wyler E, Obermayer B, M\u00fccke M, Lindberg E, Trnka F, Memczak S, Schilling M, Felkin L, Barton P, Quaife N, Vanezis K, Diecke S, Mukai M, Mah N, Oh S, Kurtz A, Schramm C, Schwinge D, Sebode M, Harakalova M, Asselbergs F, Vink A, de WR, Viswanathan S, Widjaja A, G\u00e4rtner-Rommel A, Milting H, Dos RC, Knosalla C, Mertins P, Landthaler M, Vingron M, Linke W, Seidman J, Seidman C, Rajewsky N, Ohler U, Cook S, Hubner N (2019). \u201cThe Translational Landscape of the Human Heart.\u201d Cell, 178, 242-260.e29. doi:10.1016/j.cell.2019.05.010 \u261e Link
"},{"location":"misc/publication-list/#2018","title":"2018","text":"Blanc, E., Holtgrewe, M., Dhamodaran, A., Messerschmidt, C., Willimsky, G., Blankenstein, T., Beule, D. (2018). \"Identification and Ranking of Recurrent Neo-Epitopes in Cancer\". bioRxiv. 2018/389437, 2018. doi: 10.1101/389437
Brandt, R., Uhlitz, F., Riemer, P., Giesecke, C., Schulze, S., El-Shimy, I.A., Fauler, B., Mielke, T., Mages, N., Herrmann, B.G., Sers, C., Bl\u00fcthgen, N., Morkel, M. (2018). \"Cell type-dependent differential activation of ERK by oncogenic KRAS or BRAF in the mouse intestinal epithelium\". bioRxiv. 2018/340844. doi: 10.1101/340844.
Holtgrewe, M., Knaus, A., Hildebrand, G., Pantel, J.-T., Rodriguesz de los Santos, M., Neveling, K., Goldmann, J., Schubach, M., J\u00e4ger, M., Couterier, M., Mundlos, S., Beule, D., Sperling, K., Krawitz, P. (2018). \"Multisite de novo mutations in human offspring after paternal exposure to ionizing radiation\", Nature Scientific Reports. 2018 Oct 2;8(1):14611. doi: 10.1038/s41598-018-33066-x.
Kircher M., Xiong C., Martin B, Schubach M, Inoue F, Bell R.JA., Costello J.F., Shendure J., Ahituv N. (2018). \"Saturation mutagenesis of disease-associated regulatory elements.\" bioRxiv (2018): 505362. doi: 10.1101/505362
PCAWG Transcriptome Core Group, Calabrese, C., Davidson, N.R., Fonseca1, N.A., He, Y., Kahles, A., Lehmann, K.-V., Liu, F., Shiraishi, Y., Soulette, C.M., Urban, L., Demircio\u011flu, D., Greger, L., Li, S., Liu, D., Perry, M.D., Xiang, L., Zhang, F., Zhang, J., Bailey, P., Erkek, S., Hoadley, K.A., Hou, Y., Kilpinen, H., Korbel, J.O., Marin, M.G., Markowski, J., Nandi11, T., Pan-Hammarstr\u00f6m, Q., Pedamallu, C.S., Siebert, R., Stark, S.G., Su, H., Tan, P., Waszak, S.M., Yung, C., Zhu, S., PCAWG Transcriptome Working Group, Awadalla, P., Creighton, C.J., Meyerson, M., Ouellette, B.F.F., Wu, K., Yang, H., ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Network, Brazma1, A., Brooks, A.N., G\u00f6ke, J., R\u00e4tsch, G., Schwarz, R.F., Stegle, O., Zhang, Z. (2018). \"Genomic basis for RNA alterations revealed by whole-genome analyses of 27 cancer types\". bioRxiv. 2018/183889. doi: 10.1101/183889
Guneykaya D., Ivanov A., Hernandez D.P., Haage V., Wojtas B., Meyer N., Maricos M., Jordan P., Buonfiglioli A., Gielniewski B., Ochocka N., C\u00f6mert, C., Friedrich, C., Artiles, L. S., Kaminska, B., Mertins, P., Beule, D., Kettenmann, H. (2018). \"Transcriptional and translational differences of microglia from male and female brains\", Cell reports. 2018 Sep 4;24(10):2773-83. doi: 10.1016/j.celrep.2018.08.001.
Rentzsch P, Witten D, Cooper GM, Shendure J, Kircher M. (2018). \"CADD: predicting the deleteriousness of variants throughout the human genome\", Nucleic Acids Res. 2018 Oct 29. doi: 10.1093/nar/gky1016.
Salatzki J., Foryst-Ludwig A., Bentele K., Blumrich A., Smeir E., Ban Z., Brix S., Grune J., Beyhoff N., Klopfleisch R., Dunst S., Surma, M.A., Klose, C., Rothe, M., Heinzel, F.R., Krannich, A., Kershaw, E.E., Beule, D., Schulze, P.C., Marx, N., Kintscher, U. (2018). \"Adipose tissue ATGL modifies the cardiac lipidome in pressure-overload-induced left ventricular failure\", PLoS genetics. 2018 Jan 10;14(1):e1007171. doi: 10.1371/journal.pgen.100717.
Schubach M., Re M., Robinson P.N., Valentini G. (2017) \"Imbalance-aware machine learning for predicting rare and common disease-associated non-coding variants\", Scientific reports 7:1, 2959. doi: 10.1038/s41598-017-03011-5.
Schubert M., Klinge, B., Kl\u00fcnemann M., Sieber A., Uhlitz F., Sauer S., Garnett M., Bl\u00fcthgen N., Saez-Rodriguez J. (2018). \"Perturbation-response genes reveal signaling footprints in cancer gene expression\". Nature Communications. 9: 20, 2018. doi: 10.1038/s41467-017-02391-6
"},{"location":"misc/publication-list/#2017","title":"2017","text":"Euskirchen, P., Bielle, F., Labreche, K., Kloosterman, W.P., Rosenberg, S., Daniau, M., Schmitt, C., Masliah-Planchon, J., Bourdeaut, F., Dehais, C., et al. (2017). Same-day genomic and epigenomic diagnosis of brain tumors using real-time nanopore sequencing. Acta Neuropathol 1\u201313. doi: 10.1007/s00401-017-1743-5
Euskirchen, P., Radke, J., Schmidt, M.S., Heuling, E.S., Kadikowski, E., Maricos, M., Knab, F., Grittner, U., Zerbe, N., Czabanka, M., et al. (2017). Cellular heterogeneity contributes to subtype-specific expression of ZEB1 in human glioblastoma. PLOS ONE 12, e0185376. doi: 10.1371/journal.pone.0185376
Mattei D., Ivanov A., Ferrai C., Jordan P., Guneykaya D., Buonfiglioli A., Schaafsma W., Przanowski P., Deuther-Conrad W., Brust P., Hesse S., Patt, M., Sabri, O., Ross, T.L., Eggen, B.J.L., Boddeke E.W.G.M., Kaminska, B., Beule, D., Pombo, A., Kettenmann, H., Wolf, S.A. (2017). \"Maternal immune activation results in complex microglial transcriptome signature in the adult offspring that is reversed by minocycline treatment.\" Translational psychiatry. 2017 May;7(5):e1120. doi: 10.1038/tp.2017.80.
Mamlouk, S., Childs, L. H., Aust, D., Heim, D., Melching, F., Oliveira, C., Wolf, T., Durek, P., Schumacher, D., Bl\u00e4ker, H., von Winterfeld, M., Gastl, B., M\u00f6hr, K., Menne, A., Zeugner, S., Redmer, T., Lenze, D., Tierling, S., M\u00f6bs, M., Weichert, W., Folprecht, G., Blanc, E., Beule, D., Sch\u00e4fer, R., Morkel, M., Klauschen, F., Leser, U. and Sers, C. (2017). \"DNA copy number changes define spatial patterns of heterogeneity in colorectal cancer\", Nature Communications. 2017; 8, p. 14093. doi: 10.1038/ncomms14093.
Messerschmidt, C., Holtgrewe, M. and Beule, D. (2017). \"HLA-MA: simple yet powerful matching of samples using HLA typing results\". Bioinformatics. 28, pp. 2592\u20132599. doi: 10.1093/bioinformatics/btx132.
Kammertoens, T., Friese, C., Arina, A., Idel, C., Briesemeister, D., Rothe, M., Ivanov, A., Szymborska, A., Patone, G., Kunz, S., Sommermeyer, D., Engels, B., Leisegang, M., Textor, A., Fehling, H. J., Fruttiger, M., Lohoff, M., Herrmann, A., Yu, H., Weichselbaum, R., Uckert, W., H\u00fcbner, N., Gerhardt, H., Beule, D., Schreiber, H. and Blankenstein, T. (2017). \"Tumour ischaemia by interferon-\u03b3 resembles physiological blood vessel regression\". Nature. 545(7652), pp. 98\u2013102. doi: 10.1038/nature22311.
Schulze Heuling, E., Knab, F., Radke, J., Eskilsson, E., Martinez-Ledesma, E., Koch, A., Czabanka, M., Dieterich, C., Verhaak, R.G., Harms, C., et al. (2017). Prognostic Relevance of Tumor Purity and Interaction with MGMT Methylation in Glioblastoma. Mol. Cancer Res. 15, 532\u2013540. doi: 10.1158/1541-7786.MCR-16-0322
Yaakov, G., Lerner, D., Bentele, K., Steinberger, J., Barkai, N., Bigger, J., Maisonneuve, E., Gerdes, K., Lewis, K., Dhar, N., McKinney, J. D., Gefen, O., Balaban, N. Q., Jayaraman, R., Balaban, N. Q., Merrin, J., Chait, R., Kowalik, L., Leibler, S., Balaban, N. Q., Allison, K. R., Brynildsen, M. P., Collins, J. J., Nathan, C., Lewis, K., Glickman, M. S., Sawyers, Knoechel, B., Welch, A. Z., Gibney, P. A., Botstein, D., Koshland, D. E., Levy, S. F., Ziv, N., Siegal, M. L., Stewart-Ornstein, J., Weissman, J. S., El-Samad, H., Gasch, A. P., Weinert, T., Hartwell, L., Weinert, T. A., Hartwell, L. H., Lisby, M., Rothstein, R., Mortensen, U. H., Lisby, M., Mortensen, U. H., Rothstein, R., Domkin, V., Thelander, L., Chabes, A., Hendry, J. A., Tan, G., Ou, J., Boone, C., Brown, G. W., Berry, D. B., Gasch, A. P., Lynch, M., Nishant, K. T., Serero, A., Jubin, C., Loeillet, S., Legoix-Ne, P., Nicolas, A. G., Huh, W. K., Janke, C., Lee, S. E., Blecher-Gonen, R., Martin, M., Cherry, J. M., McKenna, A., DePristo, M. A., Lawrence, M., Obenchain, V., Ye, K., Schulz, M. H., Long, Q., Apweiler, R., Ning, Z., Layer, R. M., Chiang, C., Quinlan, A. R., Hall, I. M., Faust, G. G., Hall, I. M., Boeva, V., Boeva, V., Li, H., Koren, A., Soifer, I. and Barkai, N. (2017). \"Coupling phenotypic persistence to DNA damage increases genetic diversity in severe stress\". Nature Ecology & Evolution. 1(1), pp. 497\u2013500. doi: 10.1038/s41559-016-0016.
Uhlitz, F., Sieber, A., Wyler, E., Fritsche-Guenther, R., Meisig, J., Landthaler, M., Klinger, B., Bl\u00fcthgen, N. (2017). \"An immediate-late gene expression module decodes ERK signal duration\". Molecular Systems Biology. 13: 928, 2017. doi: 10.15252/msb.20177554.
"},{"location":"misc/publication-list/#theses","title":"Theses","text":""},{"location":"misc/publication-list/#2019_1","title":"2019","text":"Schumann F. (2019). \"Establishing a pipeline for stable mutational signature detection and evaluation of variant filter effects\". Freie Universit\u00e4t Berlin. Bachelor Thesis, Bioinformatics.
"},{"location":"misc/publication-list/#2018_1","title":"2018","text":"Borgsm\u00fcller N. (2018). \"Optimization of data processing in GC-MS metabolomics\", Technische Universit\u00e4t Berlin. Master Thesis, Biotechnology.
Kuchenbecker, S.-L. (2018). \"Analysis of Antigen Receptor Repertoires Captured by High Throughput Sequencing\". Freie Universit\u00e4t Universit\u00e4t Berlin. PhD Thesis, Dr. rer. nat. URN:NBN: urn:nbn:de:kobv:188-refubium-22171-8
Schubach M. (2018). \"Learning the Non-Coding Genome\", Freie Universit\u00e4t Universit\u00e4t Berlin. PhD Thesis, Dr. rer. nat. URN:NBN: urn:nbn:de:kobv:188-refubium-23332-7
"},{"location":"misc/publication-list/#posters","title":"Posters","text":""},{"location":"misc/publication-list/#2018_2","title":"2018","text":"Roskosch, S., Hald\u00f3rsson B., Kehr, B. (2018). \"PopDel: Population-Scale Detection of Genomic Deletions\" ECCB 2018. Poster.
White T., Kehr B. (2018). \"Comprehensive extraction of structural variations from long-read DNA sequences\" WABI 2018. Poster.
"},{"location":"misc/publication-list/#2017_1","title":"2017","text":"Schubach M., Re R., Robinson P.N., Valentini G. (2017). \"Variant relevance prediction in extremely imbalanced training sets\" ISMB/ECCB 2017. Poster.
White T., Kehr B. (2017). \"Improving long-read mapping with simple lossy sequence transforms\" ISMB/ECCB 2017. Poster.
"},{"location":"ondemand/interactive/","title":"OnDemand: Interactive Sessions","text":"Interactive sessions allow you to start and manage selected apps. Depending on the app they run as servers or GUIs. Selecting My Interactive Sessions in the top menu will direct you to the overview of currently running sessions. The left-hand panel provides a short cut to start a new session of one of the provided apps.
Each running interactive session is listed. Each card corresponds to one session. The title of each card provides the name, allocated resources and the current status. Furthermore, detailed information and links are available:
Don't hit reload in your apps
Please note that the portal will use the authentication mechanisms of the apps to ensure that nobody except for you can connect to the session. This means that hitting the browsers \"reload\" button in your app will most likely not work.
Just go back to the interactive session list and click on the \"connect\" button.
"},{"location":"ondemand/interactive/#session-directories","title":"Session Directories","text":"The portal software will create a folder ondemand in your home directory. Inside, it will create session directories for each started interactive job. For technical reasons, these folders have very long names, for example:
$HOME/ondemand/data/sys/dashboard/batch_connect/sys/ood-bih-rstudio-server/output/e40e03b3-11ca-458a-855b-98e6f148c99a/This follows the pattern:
$HOME/${application name}/output/${job UUID}The job identifier used is not the Slurm job ID but an identifier internal to OnDemand. Inside this directory you will find log files and a number of scripts that are used to start your job.
If you need to debug any interactive job, start here. Also, the helpdesk will need the path to this folder to help you with interactive jobs.
You can find the name of the latest output folder with the following command:
$ ls -lhtr $HOME/${application name}/output | tail -n 1\nFor example, for RStudio Server:
$ ls -lhtr $HOME/ondemand/data/sys/dashboard/batch_connect/sys/ood-bih-rstudio-server/output | tail -n 1\nPrevent Home From Filling Up
You should probably move ~/ondemand to your work volume with the following:
$ mv ~/ondemand ~/work/ondemand\n$ ln -sr ~/work/ondemand ~/ondemand\nMake sure to delete potential interactive sessions and to logout from the Ondemand Portal first. Otherwise, the ~/ondemand folder is constantly recreated and the symlink will be just created within this folder as ~/ondemand/ondemand and thus not be used as intended.
Also, clear out ~/work/ondemand/* from time to time but take care that you don't remove the directory of any running job.
This description of starting an RStudio session is a showcase for starting other interactive apps as well.
To start the session, please go to Interactive Apps in the top menu bar and select RStudio Server or click RStudio Server in the left-hand panel.
Allocate appropriate resources and click Launch.
An info card for the RStudio Server will be added to My Interactive Sessions, and during start, it will change its state from Queued to Starting to Running. Depending on the app, resources allocated and current cluster usage, this will take a couple of seconds.
When in the final state (Running), one can directly connect to the RStudio Server to get an interactive session by clicking Connect to RStudio Server:
To use the OnDemand portal with a specific R installation including a stable set of custom packages you can use a conda enviroment from the cluster as a R source.
For this you may first need to create this conda environment including your R version of choice and all necessary packages. Specific installations of i.e. python from conda can be used similarly in other interactive apps.
channels:\n - conda-forge\n - bioconda\n - defaults\ndependencies:\n - r-base\n - r-essentials\n - r-devtools\n - bioconductor-deseq2\n - r-tidyverse\n - r-rmarkdown\n - r-knitr\n - r-dt\nSome packages (i.e. several single-cell-RNAseq analysis tools) are only available from github and not on Cran/Bioconductor. There are two ways to install such packages into a conda enviroment.
Click to expand 1) Install from inside R \\[easier option, but not pure conda\\] * First setup the conda env, ideally including all dependencies for the desired package from github (and do include r-devtools) * Then within R run `devtools::install_github('owner/repo', dependencies=F, upgrade=F, lib='/path/to/conda/env-name/lib/R/library')` * if you don't have all dependencies already installed you will have to omit dependencies=F and risk a mix of conda & native R installed packages (or just have to redo the conda env). * github_install involves a build process and still needs a bit of memory, so this might crash on the default `srun --pty bash -i` shell 2) Build packages into a local conda channel \\[takes longer, but pure conda\\]\\ This approach is mostly taken from the answers given [here](https://stackoverflow.com/questions/52061664/install-r-package-from-github-using-conda). These steps must be taken _before_ building the final env used with Rstudio * use `conda skeleton cran https://github.com/owner/repo [--git-tag vX.Y]` to generate build files * conda skeleton only works for repositories with a release/version tag. If the package you want to install does not have that, you either need to create a fork and add a such a tag, or find a fork that already did that. Downloading the code directly from github and building the package from that is also possible, but you will the need to manually set up the `meta.yaml` and `build.sh` files that conda skeleton would create. * If there is more than one release tag, do specify which one you want, it may not automatically take the most recent one. * If any r-packages from bioconductor are dependencies, conda will not find them during the build process. You will need to change the respective entries in the `meta.yaml` file created by conda skeleton. I.e. change `r-deseq2` to `bioconductor-deseq2` * Build the package with `conda build --R= [--use-local] r-` * You need to specifying the same R-version used in the final conda env * If the github package has additional dependencies from github, build those first and then add `--use-local` so the build process can find them. * The build process definitely needs more memory than the default `srun --pty bash -i` shell. It also takes quite a bit of time (much longer than installing through devtools::install_github) * Finally add the packages (+versions) you built to the environment definition (i.e. yaml file) and create the (final) conda environment. Don't forget to tell conda to use locally build packages (either supply `--use-local` or add `- local` to the channel list in the yaml file)Starting the Rstudio session via the OnDemand portal works almost as described above (see Example 1). However, you do have to select `miniforge` as R source and provide the path to your miniforge installation and (separated by a colon) the name of the (newly created) conda enviroment you want to use.
Additional notes:
.libPaths() entries and therefore a link to your previous conda installation. Creating a new project cleans .libPaths() to only the env specified in setting up the Rstudio session.Status / Stability
OnDemand Support is currently in beta phase on the BIH HPC. In case of any issues, please send an email to hpc-helpdesk@bih-charite.de.
To allow for better interactive works, BIH HPC administration has setup an Open OnDemand (OOD) portal web server.
You can find the OnDemand Portal for HPC 4 Research at:
OOD allows you to access cluster resources using a web-based graphical interface in addition to traditional SSH connections. You can then connect to jobs running graphical applications either to virtual desktops (such as Matlab) or to web apps (such as Jupyter and RStudio Server).
The following figure illustrates this.
The primary way to the cluster continues to be SSH which has several advantages. By the nature of the cluster being based on Linux servers, it will offer more features through the \"native\" access and through its lower complexity, it will offer higher stability. However, we all like to have the option of a graphical interface, at least from time to time .
The main features are:
The first prerequisite is to have a cluster account already (see Getting Access). Once you have done your first SSH connection to the cluster successfully you can start using the portal. For this you perform the following steps:
_c) then please use the \"Charit\u00e9 - Universit\u00e4tmedizin Berlin\" button, for MDC Accounts please use the \"Max Delbr\u00fcck Center Berlin\" button. Clicked the Wrong Login Button?
If you clicked the wrong button then please clear your cookies to force a logout of the system.
"},{"location":"ondemand/overview/#prepare-ondemand-folder","title":"Prepare OnDemand Folder","text":"The ondemand folder is automatically created in your home directory, and the OnDemand service searches for this folder in your home directory, i.e. it has to stay there. But as the quota in the home directory is very limited, you can easily hit the hard quota which might prevent you from working on the cluster.
To prevent this, move the ~/ondemand folder to the ~/work folder and create a symlink for the now dislocated ~/ondemand folder:
hpc-login-1:~$ mv ~/ondemand ~/work/ondemand\nhpc-login-1:~$ ln -sr ~/work/ondemand ~/ondemand\nImportant
Make sure to delete potential interactive sessions and to logout from the Ondemand Portal first. Otherwise, the ~/ondemand folder is constantly recreated and the symlink will be just created within this folder as ~/ondemand/ondemand and thus not be used as intended.
Problems with Open OnDemand?
First try to log out and login again. Next, try to clear all cookies for the domain hpc-portal.cubi.bihealth.org. Finally, try the Help > Restart Web Server link to restart the per-user nginx (PUN) server.
You will then be redirected to the dashboard screen.
Here you have access to the following actions. We will not go into detail of all of them and expect them to be self-explanatory.
Important
Please note that when using the portal then you are acting as your HPC user. Use standard best practice. Consider carefully what you do as you would from the command line (e.g., don't use the portal to browse the web from the cluster).
Outdated
This document is only valid for the old, third-generation file system and will be removed soon. Quotas of our new CephFS storage are communicated via the HPC Access web portal.
Accessing the quota report by selecting Files and then Quotas in the top menu will provide you with a detailed list of all quotas for directories that you are assigned to.
There are two types of quotas: for (a) size of and (b) number of files in a directory.
Every row in the table corresponds to a directory that you have access to. This implies your home directory (fast/users) as well as the group directory of your lab (fast/groups) and possible projects (fast/projects) (if any). Quotas are not directly implied on these directories but on the home, scratch and work subdirectories that each of subdirectory of the beforementioned directories has (for a detailed explanation see Storage and Volumes).
The following list explains the columns of the table:
/) and substituting the underscores with a slash in the (users|groups|projects)_ and _(home|scratch|work) substring. The corresponding path for name fast/users_stolpeo_c_home would be /fast/users/stolpeo_c/home.BIH HPC IT provides acess to high-performance compute (HPC) cluster systems. A cluster system bundles a high number of nodes and in the case of HPC, the focus is on performance (with contrast to high availability clusters).
"},{"location":"overview/architecture/#hpc-4-research","title":"HPC 4 Research","text":""},{"location":"overview/architecture/#cluster-hardware","title":"Cluster Hardware","text":"Users don't connect to nodes directly but rather create interactive or batch jobs to be executed by the cluster job scheduler Slurm.
As common with HPC systems, users cannot directly access the compute nodes but rather connect to so-called head nodes. The BIH HPC system provides the following head nodes:
login-1 and login-2 that accept SSH connections and are meant for low intensity, interactive work such as editing files, running screen/tmux sessions, and logging into the compute nodes. Users should run no computational tasks and no large-scale data transfer on these nodes.transfer-1 and transfer-2 also accept SSH connections. Users should run all large-scale data transfer through these nodes.After registration and client configurations, users with typically connect to the HPC system through the login nodes:
local:~$ ssh -l jdoe_c hpc-login-1.cubi.bihealth.org\nhpc-login-1:~$\nSubsequently, they might submit batch jobs to the cluster for execution through the Slurm scheduling system or open interactive sessions:
hpc-login-1:~$ sbatch job_script.sh\nhpc-login-1:~$ srun --pty bash -i\nmed0104:~$\nBIH HPC 4 Research is located in the BIH data center in Buch and connected via the BIH research network. Connections can be made from Charite, MDC, and BIH networks. The cluster is open for users with either Charite or MDC accounts after getting access through the gatekeeper proces. The system has been designed to be suitable for the processing of human genetics data from research contexts (and of course data without data privacy concerns such as public and mouse data).
"},{"location":"overview/for-the-impatient/#cluster-hardware-and-scheduling","title":"Cluster Hardware and Scheduling","text":"The cluster consists of the following major components:
hpc-login-1 and hpc-login-2 (for interactive sessions only),hpc-transfer-1 and hpc-transfer-2,hpc-cpu-{1..228}hpc-mem-{1..5},hpc-gpu-{1..7} and 1 node with 10x A40 GPUs (!) hpc-gpu-8,/fast,This is shown by the following picture:
"},{"location":"overview/for-the-impatient/#differences-between-workstations-and-clusters","title":"Differences Between Workstations and Clusters","text":"The differences include:
srun to go to a compute node.srun to go to a compute node you might end up on a different host./tmp./fast directory is shared throughout the cluster which contains your home, group home, and project directories.root or sudo permissions on the cluster.sbatch) over calling programs interactively.NB: the following might sound a bit harsh but is written with everyone's best intentions in mind (we actually like you, our user!) This addresses a lot of suboptimal (yet not dangerous, of course) points we observed in our users.
IT IS
IT IS NOT
sudo.Once logged into the cluster through the login nodes, the Slurm scheduler needs to be used to submit computing jobs. In Slurm nomenclature, cluster compute nodes are assigned to one or more partitions. Submitted jobs are assigned to nodes according to the partition's configuration.
"},{"location":"overview/job-scheduler/#partitions","title":"Partitions","text":"The BIH HPC has the partitions described below. The cluster focuses on life science applications and not \"classic HPC\" with numerical computations using MPI. Thus, all partitions except for mpi only allow to reserve resources on one node. This makes the cluster easier to use as users don't have to explicitely specify this limit when submitting their jobs.
standard","text":"Jobs are submitted to the standard partition by default. From the, the scheduler will route the jobs to their actual partition using the routing rule set described below. You can override this routing by explicitely assigning a partition (but this is discouraged).
gpu queue.highmem queue.debug, short, medium, and long long depending on their configured maximal running time. The partitions are evaluated in the order given above and the first fitting partition will be used.debug","text":"This partition is for very short jobs that should be executed quickly, e.g., for tests. The job running time is limited to one hour and at most 128 cores can be used per user but the jobs are submitted with highest priority.
debug--time 01:00:00short","text":"This partition is for jobs running only few hours. The priority of short jobs is high and many cores can be used at once to reward users for splitting their jobs into smaller parts.
short--time 04:00:00medium","text":"This partition is for jobs running for multiple days. Users can only allocate the equivalent of 4 nodes.
medium--time 7-00:00:00long","text":"This partition is for long-running tasks. Only one node can be reserved for so long to discourage really long-running jobs and encourage users for splitting their jobs into smaller parts.
long--time 14-00:00:00gpu","text":"Jobs requesting GPU resources are automatically assigned to the gpu partition.
The GPU nodes are only part of the gpu partition so they are not blocked by normal compute jobs. Maximum run time is relatively high (14 days) to allow for longer training jobs. Contact hpc-helpdesk@bih-charite.de if you have longer running jobs that you really cannot make run any shorter for assistance.
Info
Fair use rules apply. As GPU nodes are a limited resource, excessive use by single users is prohibited and can lead to mitigating actions. Be nice and cooperative with other users. Tip: getent passwd USER_NAME will give you a user's contact details.
gpu$count GPUs: -p gpu --gres=gpu:$card:$count (card=tesla or card=a40), maximum run time: --time 14-00:00:00highmem","text":"Jobs requesting more than 200 GB of RAM are automatically routed to the highmem partition.
The high memory nodes are only part of the highmem partition so they are not blocked by normal compute jobs. Maximum run time is relatively high (14 days) to allow for longer jobs. Contact hpc-helpdesk@bih-charite.de for assistance if you have longer running jobs that you really cannot make run any shorter.
Info
Fair use rules apply. As high-memory nodes are a limited resource, excessive use by single users is prohibited and can lead to mitigating actions. Be nice and cooperative with other users. Tip: getent passwd USER_NAME will give you a user's contact details.
highmem-p highmem, maximum run time: --time 14-00:00:00mpi","text":"Jobs are not routed automatically to the mpi partition but you have to explitely request the partition. This is the only partition in which more than one node can be allocated to a job.
You can submit multi-node jobs into the mpi partition. Maximum run time is relatively high (14 days) to allow for longer jobs. Don't abuse this. Contact hpc-helpdesk@bih-charite.de for assistance if you have longer running jobs that you really cannot make run any shorter.
highmem-p mpi, maximum run time: --time 14-00:00:00critical","text":"Jobs are not routed into critial automatically and the partition has to be selected manually.
This partition is for time-critical jobs with deadlines. As long as the cluster is not very busy, requests for critical jobs will be granted most of the time. However, do not use this partition without arranging with hpc-helpdesk as killing jobs will be used as the ultima ratio in case of such policy violations.
critical--time 7-00:00:00We currently provide a Grafana dashboard for monitoring various aspects of the cluster's status:
https://metrics.cubi.bihealth.org/public-dashboards/dc3e4d5b1ea049429abf39e412c47302
"},{"location":"overview/storage/","title":"Nodes and Storage Volumes","text":"No mounting on the cluster itself.
For various technical and security-related reasons it is not possible to mount anything on the cluster nodes by users. For mounting the cluster storage on your computer, please read Connecting: SSHFS Mounts.
This document gives an overview of the nodes and volumes on the cluster.
"},{"location":"overview/storage/#cluster-layout","title":"Cluster Layout","text":""},{"location":"overview/storage/#cluster-nodes","title":"Cluster Nodes","text":"The following groups of nodes are available to cluster users. There are a number of nodes that are invisible to non-admin staff, hosting the queue master and monitoring tools and providing backup storage for key critical data, but these are not shown here.
hpc-login-{1,2}hpc-login-{1,2}.cubi.bihealth.orgmed0101..0124,0127 med0133..0164med0201..0264med0301..0304med0401..0405 special purpose/high-memory machinesmed0401 and med0402med0403 and med0404med0405gpu)med0601..0616med0618..0633med0701..0764The cluster has 2.1 PB of legacy fast storage, currently available at /fast, as well as 1.6 PB of next-generation fast storage, available at /data/cephfs-1. Additionally 7.4 PB of slower \"Tier 2\" storage is available at /data/cephfs-2. Storage is provided by a Ceph storage cluster and designed for massively parallel access from an HPC system. In contrast to \"single server\" NFS systems, the system can provide large bandwidth to all cluster nodes in parallel as long as large data means relatively \"few\" files are read and written.
Storage is split into three sections:
home -- small, persistent, and safe storage, e.g., for documents and configuration files (default quota of 1 GB).work -- larger and persistent storage, e.g., for your large data files (default quota of 1 TB).scratch -- large and non-persistent storage, e.g., for temporary files, files are automatically deleted after 2 weeks (default quota of 10 TB; deletion not implemented yet).)Each user, group, and project has one or more of these sections each, e. g. for users:
/data/cephfs-1/home/users/$NAME/data/cephfs-1/home/users/$NAME/work/data/cephfs-1/home/users/$USER/scratchSee Storage and Volumes: Locations for more informatin.
"},{"location":"slurm/background/","title":"Introduction to Scheduling","text":"As explained elsewhere in more detail, an HPC cluster consists of multiple computers connected via a network and working together. Multiple users can use the system simultaneously to do their work. This means that the system needs to join multiple computers (nodes) to provide a coherent view of them and the same time partition the system to allow multiple users to work concurrently.
user 1 user 2 ...\n\n .---. .---. .---. .---.\n | J | | J | | J | | J |\n | o | | o | | o | | o | ...\n | b | | b | | b | | b |\n | 1 | | 2 | | 3 | | 4 |\n '---' '---' '---' '---'\n\n.------------------------------------------.\n| Cluster Scheduler |\n'------------------------------------------'\n\n.----------. .------------. .------------.\n| multiple | | separate | | computers |\n'----------' '------------' '------------'\nOverall, this partitioning is not so different from how your workstation or laptop works. Most likely, your computer (or even your smartphone) has multiple processors (or cores). You can run multiple programs on the same computer and the fact that (a) there is more than one core and (b) there is more than one program running is not known to the running programs (unless they explicitly communicate with each other). Different programs can explicitly take advantage of the multiple processor cores. The main difference is that you normally use your computer in an interactive fashion (you perform an action and expect an immediate reaction).
Even with a single processor (and core), your computer manages to run more than one program at the same time. This is done with the so-called time-slicing approach where the operating system lets each programs run in turn for a short time (a few milliseconds). A program with a higher priority will get more time slices than one with a lower (e.g., your audio player has real-time requirements and you will hear artifacts if it is starved for compute resources). Your operating system protects programs from each other by creating an address space for each. When two programs are running, the value of the memory at any given position in one program is independent from the value in the other program. Your operating system offers explicit functionality for sharing certain memory areas that two programs can use to exchange data efficiently.
Similarly, file permissions with Unix users/groups or Unix/Windows ACLs (access control lists) are used to isolate users from each other. Programs can share data by accessing the same file if they can both access it. There are special files called sockets that allow for network-like inter-process communication but of course two programs on the same computer can also connect (virtually) via the computer network (no data will actually go through a cable).
"},{"location":"slurm/background/#interlude-resource-types","title":"Interlude: Resource Types","text":"As another diversion, let us consider how Unix manages its resources. This is important to understand when requesting resources from the scheduler later on.
First of all, a computer might offer a certain feature such as a specific hardware platform or special network connection. Examples for this on the BIH HPC are specific Intel processor generations such as haswell or the availability of Infiniband networking. You can request these with so-called constraints; they are not allocated to specific jobs.
Second, there are resources that are allocated to specific jobs. The most important resources here are:
Generally, once a resource has been allocated to one job, it is not available to another. This means if you allocating more resources to your job that you actually need (overallocation) then those resources are not available to other jobs (whether they are your jobs or those of other users). This will be explained further below.
Another example of resource allocation are licenses. The BIH HPC has a few Matlab 2016b licenses that users can request. As long as a license is allocated to one job, it is unavailable to another.
"},{"location":"slurm/background/#nodes-sockets-processors-cores-threads","title":"Nodes, Sockets, Processors, Cores, Threads","text":"Regarding compute resources, Slurm differentiates between:
In most cases, you will use one compute node only. When using more than one node, you will need to use some form of message passing, e.g., MPI, so processes on different nodes can communicate. On a single node you would mostly use single- or multi-threaded processes, or multiple processes.
Above: Slurm's nomenclature for sockets, processors, cores, and threads (from Slurm Documentation).
Co-locating processes/threads on the same socket has certain implications that are mostly useful for numerical applications. We will not further go into detail here. Slurm provides many different features of ways to specify allocation of \"pinning\" to specific process locations. If you need this feature, we trust that you find sufficient explanation in the Slurm documentation.
Usually, you would allocate multiple cores (a term Slurm uses synonymously with processors) on a single node (allocation on a single node is the default).
"},{"location":"slurm/background/#how-scheduling-works","title":"How Scheduling Works","text":"Slurm is an acronym for \"Simple Linux Unix Resource Manager\" (note that the word \"scheduler\" does not occur here). Actually, one classically differentiates between the managing of resources and the scheduling of jobs that use them. The resource manager allocates resources according to a user's request for a job and ensures that there are no conflicts. If the required resources are not available, the scheduler puts the user's job into a queue. Later, when then requested resources become available the scheduler assigns them to the job and runs it. In the following, both resource allocation and the running of the job are described as being done by the scheduler.
The interesting case occurs when there are not enough resources available for at least two jobs submitted to the scheduler. The scheduler has to decide how to proceed. Consider the simplified case of only scheduling cores. Each job will request a number of cores. The scheduler will then generate a scheduling plan that might look as follows.
core\n ^\n4 | |---job2---|\n3 | |---job2---|\n2 | |---job2---|\n1 | |--job1--|\n +--------------------------> t time\n 5 1 1 2\n 0 5 0\njob1 has been allocated one core and job2 has been allocated two cores. When job3, requesting one core is submitted at t = 5, it has to wait at least as long until job1 is finished. If job3 requested two or more cores, it would have to wait at least until job2 also finished.
We can now ask several questions, including the following:
Also see the Slurm Frequently Asked Questions.
Please note that even if all jobs were known at the start of time, scheduling is still a so-called NP-complete problem. Entire computer science journals and books are dedicated only to scheduling. Things get more complex in the case of online scheduling, in which new jobs can appear at any time. In practice, Slurm does a fantastic job with its heuristics but it heavily relies on parameter tuning. HPC administration is constantly working on optimizing the scheduler settings. Note that you can use the --format option to the squeue command to request that it shows you information about job scheduling (in particular, see the %S field, which will show you the expected start time for a job, assuming Slurm has calculated it). See man squeue for details. If you observe inexplicable behavior, please notify us at hpc-helpdesk@bih-charite.de.
In Slurm, the nodes of a cluster are split into partitions. Nodes are assigned to one or more partition (see the Job Scheduler section for details). Jobs can also be assigned to one or more partitions and are executed on nodes of the given partition.
In the BIH HPC, partitions are used to stratify jobs of certain running times and to provide different quality of service (e.g., maximal number of CPU cores available to a user for jobs of a certain running time and size). The partitions gpu and highmem provide special hardware (the nodes are not assigned to other partitions) and the mpi partition allows MPI-parallelism and the allocation of jobs to more than one node. The Job Scheduler provides further details.
This page contains assorted Slurm commands and Bash snippets that should be helpful.
man pages!
$ man sinfo\n$ man scontrol\n$ man squeue\n# etc...\ninteractive sessions
hpc-login-1:~$ srun --pty bash\nmed0740:~$ echo \"Hello World\"\nmed0740:~$ exit\nbatch submission
hpc-login-1:~$ sbatch script.sh\nSubmitted batch job 2\nhpc-login-1:~$ squeue\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 27 debug script.s holtgrem R 0:06 1 med0703\nlisting nodes
$ sinfo -N\nNODELIST NODES PARTITION STATE\nmed0740 1 debug* idle\nmed0741 1 debug* down*\nmed0742 1 debug* down*\n\n$ scontrol show nodes\nNodeName=med0740 Arch=x86_64 CoresPerSocket=8\n CPUAlloc=0 CPUTot=32 CPULoad=0.06\n AvailableFeatures=(null)\n[...]\n\n$ scontrol show nodes med0740\nNodeName=med0740 Arch=x86_64 CoresPerSocket=8\n CPUAlloc=0 CPUTot=32 CPULoad=0.06\n AvailableFeatures=(null)\n ActiveFeatures=(null)\n Gres=(null)\n NodeAddr=med0740 NodeHostName=med0740 Version=20.02.0\n OS=Linux 3.10.0-1062.12.1.el7.x86_64 #1 SMP Tue Feb 4 23:02:59 UTC 2020\n RealMemory=1 AllocMem=0 FreeMem=174388 Sockets=2 Boards=1\n State=IDLE ThreadsPerCore=2 TmpDisk=0 Weight=1 Owner=N/A MCS_label=N/A\n Partitions=debug\n BootTime=2020-03-05T00:54:15 SlurmdStartTime=2020-03-05T16:23:25\n CfgTRES=cpu=32,mem=1M,billing=32\n AllocTRES=\n CapWatts=n/a\n CurrentWatts=0 AveWatts=0\n ExtSensorsJoules=n/s ExtSensorsWatts=0 ExtSensorsTemp=n/s\nqueue states
$ squeue\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n$ squeue -u holtgrem_c\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\nnode resources
$ sinfo -o \"%20N %10c %10m %25f %10G \"\nadditional resources such as GPUs
$ sinfo -o \"%N %G\"\nlisting job details
$ scontrol show job 225\nJobId=225 JobName=bash\n UserId=XXX(135001) GroupId=XXX(30069) MCS_label=N/A\n Priority=4294901580 Nice=0 Account=(null) QOS=normal\n JobState=FAILED Reason=NonZeroExitCode Dependency=(null)\n Requeue=1 Restarts=0 BatchFlag=0 Reboot=0 ExitCode=130:0\n RunTime=00:16:27 TimeLimit=14-00:00:00 TimeMin=N/A\n SubmitTime=2020-03-23T11:34:26 EligibleTime=2020-03-23T11:34:26\n AccrueTime=Unknown\n StartTime=2020-03-23T11:34:26 EndTime=2020-03-23T11:50:53 Deadline=N/A\n SuspendTime=None SecsPreSuspend=0 LastSchedEval=2020-03-23T11:34:26\n Partition=gpu AllocNode:Sid=hpc-login-1:1918\n ReqNodeList=(null) ExcNodeList=(null)\n NodeList=med0301\n BatchHost=med0301\n NumNodes=1 NumCPUs=2 NumTasks=0 CPUs/Task=1 ReqB:S:C:T=0:0:*:*\n TRES=cpu=2,node=1,billing=2\n Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*\n MinCPUsNode=1 MinMemoryNode=0 MinTmpDiskNode=0\n Features=(null) DelayBoot=00:00:00\n OverSubscribe=OK Contiguous=0 Licenses=(null) Network=(null)\n Command=bash\n WorkDir=XXX\n Power=\n TresPerNode=gpu:tesla:4\n MailUser=(null) MailType=NONE\nhost:~$ squeue\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) \n 1177 medium bash jweiner_ R 4-21:52:24 1 med0127 \n 1192 medium bash jweiner_ R 4-07:08:40 1 med0127 \n 1209 highmem bash mkuhrin_ R 2-01:07:17 1 med0402 \n 1210 gpu bash hilberta R 1-10:30:34 1 med0304 \n 1213 long bash schubacm R 1-09:42:27 1 med0127 \n 2401 gpu bash ramkem_c R 1-05:14:53 1 med0303 \n 2431 medium ngs_mapp holtgrem R 1-05:01:41 1 med0127 \n 2437 critical snakejob holtgrem R 1-05:01:34 1 med0135 \n 2733 debug bash schubacm R 7:36:42 1 med0127 \n 3029 critical ngs_mapp holtgrem R 5:59:07 1 med0127 \n 3030 critical snakejob holtgrem R 5:56:23 1 med0134 \n 3031 critical snakejob holtgrem R 5:56:23 1 med0137 \n 3032 critical snakejob holtgrem R 5:56:23 1 med0137 \n 3033 critical snakejob holtgrem R 5:56:23 1 med0138 \n 3034 critical snakejob holtgrem R 5:56:23 1 med0138 \n 3035 critical snakejob holtgrem R 5:56:20 1 med0139 \n 3036 critical snakejob holtgrem R 5:56:20 1 med0139 \n 3037 critical snakejob holtgrem R 5:56:20 1 med0140 \n 3038 critical snakejob holtgrem R 5:56:20 1 med0140 \n 3039 critical snakejob holtgrem R 5:56:20 1 med0141 \n 3040 critical snakejob holtgrem R 5:56:20 1 med0141 \n 3041 critical snakejob holtgrem R 5:56:20 1 med0142 \n 3042 critical snakejob holtgrem R 5:56:20 1 med0142 \n 3043 critical snakejob holtgrem R 5:56:20 1 med0143 \n 3044 critical snakejob holtgrem R 5:56:20 1 med0143 \n 3063 long bash schubacm R 4:12:37 1 med0127 \n 3066 long bash schubacm R 4:11:47 1 med0127 \n 3113 medium ngs_mapp holtgrem R 1:52:33 1 med0708 \n 3118 medium snakejob holtgrem R 1:50:38 1 med0133 \n 3119 medium snakejob holtgrem R 1:50:38 1 med0703 \n 3126 medium snakejob holtgrem R 1:50:38 1 med0706 \n 3127 medium snakejob holtgrem R 1:50:38 1 med0144 \n 3128 medium snakejob holtgrem R 1:50:38 1 med0144 \n 3133 medium snakejob holtgrem R 1:50:35 1 med0147 \n 3134 medium snakejob holtgrem R 1:50:35 1 med0147 \n 3135 medium snakejob holtgrem R 1:50:35 1 med0148 \n 3136 medium snakejob holtgrem R 1:50:35 1 med0148 \n 3138 medium snakejob holtgrem R 1:50:35 1 med0104 \nhost:~$ squeue -o \"%.10i %9P %20j %10u %.2t %.10M %.6D %10R %b\"\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(R TRES_PER_NODE\n 1177 medium bash jweiner_m R 4-21:52:22 1 med0127 N/A\n 1192 medium bash jweiner_m R 4-07:08:38 1 med0127 N/A\n 1209 highmem bash mkuhrin_m R 2-01:07:15 1 med0402 N/A\n 1210 gpu bash hilberta_c R 1-10:30:32 1 med0304 gpu:tesla:4\n 1213 long bash schubacm_c R 1-09:42:25 1 med0127 N/A\n 2401 gpu bash ramkem_c R 1-05:14:51 1 med0303 gpu:tesla:1\n 2431 medium ngs_mapping holtgrem_c R 1-05:01:39 1 med0127 N/A\n 2437 critical snakejob.ngs_mapping holtgrem_c R 1-05:01:32 1 med0135 N/A\n 2733 debug bash schubacm_c R 7:36:40 1 med0127 N/A\n 3029 critical ngs_mapping holtgrem_c R 5:59:05 1 med0127 N/A\n 3030 critical snakejob.ngs_mapping holtgrem_c R 5:56:21 1 med0134 N/A\n 3031 critical snakejob.ngs_mapping holtgrem_c R 5:56:21 1 med0137 N/A\n 3032 critical snakejob.ngs_mapping holtgrem_c R 5:56:21 1 med0137 N/A\n 3033 critical snakejob.ngs_mapping holtgrem_c R 5:56:21 1 med0138 N/A\n 3034 critical snakejob.ngs_mapping holtgrem_c R 5:56:21 1 med0138 N/A\n 3035 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0139 N/A\n 3036 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0139 N/A\n 3037 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0140 N/A\n 3038 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0140 N/A\n 3039 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0141 N/A\n 3040 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0141 N/A\n 3041 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0142 N/A\n 3042 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0142 N/A\n 3043 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0143 N/A\n 3044 critical snakejob.ngs_mapping holtgrem_c R 5:56:18 1 med0143 N/A\n 3063 long bash schubacm_c R 4:12:35 1 med0127 N/A\n 3066 long bash schubacm_c R 4:11:45 1 med0127 N/A\n 3113 medium ngs_mapping holtgrem_c R 1:52:31 1 med0708 N/A\n 3118 medium snakejob.ngs_mapping holtgrem_c R 1:50:36 1 med0133 N/A\n 3119 medium snakejob.ngs_mapping holtgrem_c R 1:50:36 1 med0703 N/A\n 3126 medium snakejob.ngs_mapping holtgrem_c R 1:50:36 1 med0706 N/A\n 3127 medium snakejob.ngs_mapping holtgrem_c R 1:50:36 1 med0144 N/A\n 3128 medium snakejob.ngs_mapping holtgrem_c R 1:50:36 1 med0144 N/A\n 3133 medium snakejob.ngs_mapping holtgrem_c R 1:50:33 1 med0147 N/A\n 3134 medium snakejob.ngs_mapping holtgrem_c R 1:50:33 1 med0147 N/A\n 3135 medium snakejob.ngs_mapping holtgrem_c R 1:50:33 1 med0148 N/A\n 3136 medium snakejob.ngs_mapping holtgrem_c R 1:50:33 1 med0148 N/A\n 3138 medium snakejob.ngs_mapping holtgrem_c R 1:50:33 1 med0104 N/A\nhost:~$ sinfo\nPARTITION AVAIL TIMELIMIT NODES STATE NODELIST \ndebug* up 8:00:00 11 drain med[0707,0709-0710,0740-0742,0744-0745,0749,0752,0755] \ndebug* up 8:00:00 8 mix med[0104,0127,0133-0135,0703,0706,0708] \ndebug* up 8:00:00 10 alloc med[0137-0144,0147-0148] \ndebug* up 8:00:00 103 idle med[0105-0124,0136,0145-0146,0151-0164,0201-0264,0704-0705] \nmedium up 7-00:00:00 11 drain med[0707,0709-0710,0740-0742,0744-0745,0749,0752,0755] \nmedium up 7-00:00:00 8 mix med[0104,0127,0133-0135,0703,0706,0708] \nmedium up 7-00:00:00 10 alloc med[0137-0144,0147-0148] \nmedium up 7-00:00:00 103 idle med[0105-0124,0136,0145-0146,0151-0164,0201-0264,0704-0705] \nlong up 28-00:00:0 11 drain med[0707,0709-0710,0740-0742,0744-0745,0749,0752,0755] \nlong up 28-00:00:0 8 mix med[0104,0127,0133-0135,0703,0706,0708] \nlong up 28-00:00:0 10 alloc med[0137-0144,0147-0148] \nlong up 28-00:00:0 103 idle med[0105-0124,0136,0145-0146,0151-0164,0201-0264,0704-0705] \ncritical up 7-00:00:00 11 drain med[0707,0709-0710,0740-0742,0744-0745,0749,0752,0755] \ncritical up 7-00:00:00 8 mix med[0104,0127,0133-0135,0703,0706,0708] \ncritical up 7-00:00:00 10 alloc med[0137-0144,0147-0148] \ncritical up 7-00:00:00 103 idle med[0105-0124,0136,0145-0146,0151-0164,0201-0264,0704-0705] \nhighmem up 14-00:00:0 1 mix med0402 \nhighmem up 14-00:00:0 3 idle med[0401,0403-0404] \ngpu up 14-00:00:0 2 mix med[0303-0304] \ngpu up 14-00:00:0 2 idle med[0301-0302] \nsacct","text":"Perform queries to the Slurm accounting information.
Representative Example
hpc-login-1:~$ sacct -j 1607103\n JobID JobName Partition Account AllocCPUS State ExitCode\n------------ ---------- ---------- ---------- ---------- ---------- --------\n1607103 wgs_sv_an+ medium 1 PENDING 0:0\nThe sacct command displays information from the Slurm accounting service. The Slurm scheduler only knows about active or completing (very recently active) jobs. The accouting system also knows about currently running jobs so it is the more robust way to query information about jobs. However, not all information is available to the accouting system, so scontrol show job and squeue provide more information about current and pending jbos.
Slurm Documentation: sacct
Please also see the official Slurm documentation on sacct.
"},{"location":"slurm/commands-sacct/#important-arguments","title":"Important Arguments","text":"Also see all important arguments of the sbatch command.
--jobs -- The job(s) to query for.--format -- Define attributes to retrieve.--long -- Get a lot of information from the database, consider to pipe into | less -S.sacct over scontrol and squeue.sattach","text":"The sattach command allows you to connect the standard input, output, and error streams to your current terminals ession.
Representative Example
hpc-login-1:~$ sattach 12345.0\n[...output of your job...]\nmed0211:~$ [Ctrl-C]\nhpc-login-1:~$\nPress Ctrl-C to detach from the current session. Please note that you will have to give the job ID as well as step step ID. For most cases, simply append \".0\" to your job ID.
Slurm Documentation: sattach
Please also see the official Slurm documentation on srun.
"},{"location":"slurm/commands-sattach/#important-arguments","title":"Important Arguments","text":"--pty -- Execute task zero in pseudo terminal.--verbose -- Increase verbosity of sattach.sbatch","text":"The sbatch command allows you to put a job into the scheduler's queue to be executed at a later time.
Representative Example
# Execute job.sh in partition medium with 4 threads and 4GB of RAM total for a\n# running time of up to one day.\nhpc-login-1:~$ sbatch --partition=medium --mem=4G --ntasks 4 --time=1-00 job.sh\nSubmitted batch job JOB_ID\nThe command will create a batch job and add it to the queue to be executed at a later point in time.
Slurm Documentation: sbatch
Please also see the official Slurm documentation on sbatch.
"},{"location":"slurm/commands-sbatch/#important-arguments","title":"Important Arguments","text":"--array -- Submit jobs as array jobs. Also see the section [#array-jobs] below.--nodes -- The number of nodes to allocate. This is only given here as an important argument as the maximum number of nodes allocatable to any partition but mpi is set to one (1). This is done as there are few users on the BIH HPC that actually use multi-node paralleilsm. Rather, most users will use multi-core parallelism and might forget to limit the number of nodes which causes inefficient allocation of resources.--cpus-per-task -- This corresponds to the number of CPU cores allocated to each task.--mem -- The memory to allocate for the job. As you can define minimal and maximal number of tasks/CPUs/cores, you could also specify --mem-per-cpu and get more flexible scheduling of your job.--gres -- Generic resource allocation. On the BIH HPC, this is only used for allocating GPUS, e.g., with --gres=gpu:tesla:2, a user could allocate two NVIDIA Tesla GPUs on the same host (use a40 instead of tesla for the A40 GPUs).--licenses -- On the BIH HPC, this is used for the allocation of MATLAB 2016b licenses only.--partition -- The partition to run in. Also see the Job Scheduler section.--time -- Specify the running time, see man sbatch or the official Slurm documentation on srun for supported formats. **Please note that the DRMA API only accepts the hours:minutes format.--dependency -- Specify dependencies on other jobs, e.g., using --dependency afterok:JOBID to only execute if the job with ID JOBID finished successfully or --dependency after:JOBID to wait for a job to finish regardless of its termination status.--constraint -- Require one or more features from your node. On the BIH HPC, the processor generation is defined as a feature on the nodes, e.g., haswell, or special networking such as infiniband. You can have a look at /etc/slurm/slurm.conf on all configured features.--output -- The path to the output log file (by default joining stdout and stderr, see the man page on --error on how to redirect stderr separately). A various number of placeholders is available, see the \"filename pattern\" section of man sbatch or the official Slurm documentation on srun.--mail-type=<type> -- Send out notifications by email when an event occurs. Use FAIL to get emails when your job fails. Also see the documentation of sbatch in the Slurm manual.--mail-user=<email> -- The email address to send to. Must end in @charite.de, @mdc-berlin.de, or @bih-charite.de.Ensure your --output directory exists!
In the case that the path to the log/output file does not exist, the job will just fail. scontrol show job ID will report JobState=FAILED Reason=NonZeroExitCode. Regrettably, no further information is displayed to you as the user. Always check that the path to the directories in StdErr and StdOut exists when checking scontrol show job ID.
--job-nameAlso see the section Slurm Job Scripts on how to embed the sbatch parameters in #SBATCH lines.
If you have many (say, more than 10) similar jobs (e.g., when performing a grid search), you can also use array jobs. However, you should also consider whether it would make sense to increase the time of your jobs, e.g, to be at least ~10min.
You can submit array jobs by specifying -a EXPR or --array EXPR where EXPR is a range or a list (of course, you can also add this as an #SBATCH header in your job script). For example:
hpc-login-1 ~# sbatch -a 1-3 grid_search.sh\nhpc-login-1 ~# sbatch -a 1,2,5-10 grid_search.sh\nThis will submit grid_search.sh with certain variables set:
SLURM_ARRAY_JOB_ID -- the ID of the first jobSLURM_ARRAY_TASK_ID -- the index of the job in the arraySLURM_ARRAY_TASK_COUNT -- number of submitted jobs in arraySLURM_ARRAY_TASK_MAX -- higehst job array index valueSLURM_ARRAY_TASK_MIN -- lowest job array index valueUsing array jobs has several advantages:
Also see Slurm documentation on job arrays.
For example, if you submit sbatch --array=1-3 grid_search.sh and slurm responsds with Submitted batch job 36 then the script will be run three times with the following prameters set:
SLURM_JOB_ID=36\nSLURM_ARRAY_JOB_ID=36\nSLURM_ARRAY_TASK_ID=1\nSLURM_ARRAY_TASK_COUNT=3\nSLURM_ARRAY_TASK_MAX=3\nSLURM_ARRAY_TASK_MIN=1\n\nSLURM_JOB_ID=37\nSLURM_ARRAY_JOB_ID=36\nSLURM_ARRAY_TASK_ID=2\nSLURM_ARRAY_TASK_COUNT=3\nSLURM_ARRAY_TASK_MAX=3\nSLURM_ARRAY_TASK_MIN=1\n\nSLURM_JOB_ID=38\nSLURM_ARRAY_JOB_ID=36\nSLURM_ARRAY_TASK_ID=3\nSLURM_ARRAY_TASK_COUNT=3\nSLURM_ARRAY_TASK_MAX=3\nSLURM_ARRAY_TASK_MIN=1\nsbatch are governed by resource allocations, in particular:sbatch jobs have a maximal running time set,sbatch jobs have a maximal memory and number of cores set, andscontrol show job JOBID.scancel","text":"Terminate a running Slurm job.
Representative Example
hpc-login-1:~$ scancel 1703828\nhpc-login-1:~$\nThis command allows to terminate one or more running jobs (of course, non-superusers can only terminate their own jobs).
Slurm Documentation: scancel
Please also see the official Slurm documentation on srun.
"},{"location":"slurm/commands-scontrol/","title":"Slurm Command:scontrol","text":"The scontrol allows to query detailed information from the scheduler and perform manipulation. Object manipulation is less important for normal users.
Representative Example
hpc-login-1:~$ scontrol show job 1607103\nJobId=1607103 JobName=wgs_sv_annotation\n UserId=holtgrem_c(100131) GroupId=hpc-ag-cubi(5272) MCS_label=N/A\n Priority=748 Nice=0 Account=(null) QOS=normal\n [...]\nhpc-login-1:~$ scontrol show node med02[01-32]\nNodeName=med0201 Arch=x86_64 CoresPerSocket=8\n CPUAlloc=0 CPUTot=32 CPULoad=0.01\n AvailableFeatures=ivybridge,infiniband\n ActiveFeatures=ivybridge,infiniband\n [...]\nhpc-login-1:~$ scontrol show partition medium\nPartitionName=medium\n AllowGroups=ALL AllowAccounts=ALL AllowQos=ALL\n AllocNodes=ALL Default=NO QoS=medium\n DefaultTime=NONE DisableRootJobs=NO ExclusiveUser=NO GraceTime=0 Hidden=NO\n [...]\nThis command allows to query all information for an object from Slurm, e.g., jobs, nodes, or partitions. The command also accepts ranges of jobs and hosts. It is most useful to get the information of one or a few objects from the scheduler.
Slurm Documentation: scontrol
Please also see the official Slurm documentation on scontrol.
"},{"location":"slurm/commands-scontrol/#important-sub-commands","title":"Important Sub commands","text":"scontrol show job -- Show details on jobs.scontrol show partition -- Show details on partitions.scontrol show node -- Show details on nodes.scontrol help -- Show help.scontrol -- Start an interactive scontrol shell / REPL (read-eval-print loop).scontrol can only work on jobs that are pending (in the queue), running, or in \"completing' state.sacct command.sinfo","text":"The sinfo command allows you to query the current cluster status.
Representative Example
hpc-login-1:~$ sinfo\nPARTITION AVAIL TIMELIMIT NODES STATE NODELIST\n[...]\nmedium up 7-00:00:00 10 drain* med[0101-0103,0125-0126,0128-0132]\nmedium up 7-00:00:00 1 down* med0243\nmedium up 7-00:00:00 31 mix med[0104,0106-0122,0124,0133,0232-0233,0237-0238,0241-0242,0244,0263-0264,0503,0506]\nmedium up 7-00:00:00 5 alloc med[0105,0123,0127,0239-0240]\nmedium up 7-00:00:00 193 idle med[0134-0164,0201-0231,0234-0236,0245-0262,0501-0502,0504-0505,0507-0516,0601-0632,0701-0764]\n[...]\nhpc-login-1:$ sinfo --summarize\nPARTITION AVAIL TIMELIMIT NODES(A/I/O/T) NODELIST\ndebug* up 8:00:00 38/191/11/240 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\nmedium up 7-00:00:00 38/191/11/240 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\nlong up 28-00:00:0 38/191/11/240 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\ncritical up 7-00:00:00 25/141/10/176 med[0101-0164,0501-0516,0601-0632,0701-0764]\nhighmem up 14-00:00:0 1/2/1/4 med[0401-0404]\ngpu up 14-00:00:0 3/0/1/4 med[0301-0304]\nmpi up 14-00:00:0 38/191/11/240 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\nThis command will summaries the state of nodes by different criteria (e.g., by partition or globally).
Slurm Documentation: sinfo
Please also see the official Slurm documentation on srun.
"},{"location":"slurm/commands-sinfo/#important-arguments","title":"Important Arguments","text":"Also see all important arguments of the sinfo command.
--summarize -- Summarize the node state by partition.--nodes -- Select the nodes to show the status for, e.g., display the status of all GPU nodes with sinfo -n med030[1-4].The most important node states are:
down -- node is marked as offlinedraining -- node will not accept any more jobs but has jobs running on itdrained -- node will not accept any more jobs and has no jobs running on it, but is not offline yetidle -- node is ready to run jobsallocated -- node is fully allocated (e.g., CPU, RAM, or GPU limit has been reached)mixed -- node is running jobs but there is space for moresqueue","text":"The squeue command allows you to view currently running and pending jobs.
Representative Example
hpc-login-1:~$ squeue\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 1583165 highmem 20200702 usr PD 0:00 1 (DependencyNeverSatisfied)\n 1605901 critical variant_ holtgrem PD 0:00 1 (DependencyNeverSatisfied)\n 1605902 critical variant_ holtgrem PD 0:00 1 (Dependency)\n 1605905 critical variant_ holtgrem PD 0:00 1 (DependencyNeverSatisfied)\n 1605916 critical wgs_sv_c holtgrem PD 0:00 1 (Dependency)\n 1607103 medium wgs_sv_a holtgrem PD 0:00 1 (DependencyNeverSatisfied)\n[...]\nSlurm Documentation: squeue
Please also see the official Slurm documentation on squeue.
"},{"location":"slurm/commands-squeue/#important-arguments","title":"Important Arguments","text":"--nodelist -- Only display jobs running on certain nodes (e.g., GPU nodes).--format -- Define the format to print, see man squeue for details. See below for a format string that includes the jobid, partition, job name, user name, job status, running time, number of nodes, number of CPU cores, and allocated GPUs.The following aliases in ~/.bashrc will allow you to print a long and informative squeue output with sq, pipe it into less with sql, get only your jobs (adjust the alias to your account) using sqme and pipe that into less with sqmel.
alias sq='squeue -o \"%.10i %9P %60j %10u %.2t %.10M %.6D %.4C %10R %b\" \"$@\"'\nalias sql='sq \"$@\" | less -S'\nalias sqme='sq -u YOURUSER_c_or_m \"$@\"'\nalias sqmel='sqme \"$@\" | less -S'\nsrun","text":"The srun command allows you to run a command now.
Representative Example
hpc-login-1:~$ srun --pty bash -i\nmed0201:~$\nThe command will perform a resource allocation with the scheduler (and wait until it has allocated the requested resources) first. Most importantly, you can specify the --pty argument which will connect the current terminal's standard output, error, and input to your current one. This allows you to run interactive jobs such as shells with srun --pty bash -i.
Slurm Documentation: srun
Please also see the official Slurm documentation on srun.
"},{"location":"slurm/commands-srun/#important-arguments","title":"Important Arguments","text":"Also see all important arguments of the sbatch command.
--pty -- Connect current terminal to the job's stdoud/stderr/stdin.--x11 -- Setup X11 forwarding.--immediate -- Immediately terminate if the resources to run the job are not available, do not wait.--test-only -- Don't run anything, but only estimate when the job would be scheduled.srun are governed by resource allocations, in particular:srun jobs have a maximal running time set,srun jobs have a maximal memory and number of cores set, andscontrol show job JOBID.In the sections Slurm Quickstart and Slurm Cheat Sheet, we have seen that sinfo and squeue allow for the compact display partitions/nodes and node information. In contrast, scontrol show job <id> and scontrol show partition <id> and scontrol show node <id> show comprehensive information that quickly gets hard to comprehend for multiple entries.
Now you might ask: is there anything in between? And: yes, there is.
You can tune the output of sinfo and squeue using parameters, in particular by providing format strings. All of this is described in the man pages of the commands that you can display with man sinfo and man squeue on the cluster.
sinfo Output","text":"Notable arguments of sinfo are:
-N, --Node -- uncompress the usual lines and display one line per node and partition.-s, --summarize -- compress the node state, more compact display.-R, --list-reasons -- for nodes that are not up, display reason string provided by admin.-o <fmt>, --format=<fmt> -- use format string for display.The most interesting argument is -o/--format. The man page lists the following values that are used when using other arguments. In other words, many of the display modifications could also be applied with -o/--format.
default \"%#P %.5a %.10l %.6D %.6t %N\"\n--summarize \"%#P %.5a %.10l %.16F %N\"\n--long \"%#P %.5a %.10l %.10s %.4r %.8h %.10g %.6D %.11T %N\"\n--Node \"%#N %.6D %#P %6t\"\n--long --Node \"%#N %.6D %#P %.11T %.4c %.8z %.6m %.8d %.6w %.8f %20E\"\n--list-reasons \"%20E %9u %19H %N\"\n--long --list-reasons\n \"%20E %12U %19H %6t %N\"\nThe best way to learn more about this is to play around with sinfo -o, starting out with one of the format strings above. Details about the format strings are described in man sinfo. Some remarks here:
%<num><char> displays the value represented by <char> padded with spaces to the right such that a width of <num> is reached,%.<num><char> displays the value represented by <char> padded with spaces to the left such that a width of <num> is reached, and%#<char> displays the value represented by <char> padded with spaces to the max length of the value represented by <char> (this is a \"virtual\" value, used internally only, you cannot use this and you will have to place an integer here).For example, to create a grouped display with reasons for being down use:
hpc-login-1:~$ sinfo -o \"%10P %.5a %.10l %.16F %40N %E\"\nPARTITION AVAIL TIMELIMIT NODES(A/I/O/T) NODELIST REASON\ndebug* up 8:00:00 0/0/16/16 med[0703-0710,0740-0742,0744-0745,0749,0 bogus node\ndebug* up 8:00:00 18/98/0/116 med[0104-0124,0127,0133-0148,0151-0164,0 none\nmedium up 7-00:00:00 0/0/16/16 med[0703-0710,0740-0742,0744-0745,0749,0 bogus node\nmedium up 7-00:00:00 18/98/0/116 med[0104-0124,0127,0133-0148,0151-0164,0 none\nlong up 28-00:00:0 0/0/16/16 med[0703-0710,0740-0742,0744-0745,0749,0 bogus node\nlong up 28-00:00:0 18/98/0/116 med[0104-0124,0127,0133-0148,0151-0164,0 none\ncritical up 7-00:00:00 0/0/16/16 med[0703-0710,0740-0742,0744-0745,0749,0 bogus node\ncritical up 7-00:00:00 18/98/0/116 med[0104-0124,0127,0133-0148,0151-0164,0 none\nhighmem up 14-00:00:0 0/4/0/4 med[0401-0404] none\ngpu up 14-00:00:0 3/1/0/4 med[0301-0304] none\nsqueue Output","text":"The standard squeue output might yield the following
hpc-login-1:~$ squeue | head\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 3149 medium variant_ holtgrem PD 0:00 1 (Dependency)\n 1177 medium bash jweiner_ R 6-03:32:41 1 med0127\n 1192 medium bash jweiner_ R 5-12:48:57 1 med0127\n 1210 gpu bash hilberta R 2-16:10:51 1 med0304\n 1213 long bash schubacm R 2-15:22:44 1 med0127\n 2401 gpu bash ramkem_c R 2-10:55:10 1 med0303\n 3063 long bash schubacm R 1-09:52:54 1 med0127\n 3066 long bash schubacm R 1-09:52:04 1 med0127\n 3147 medium ngs_mapp holtgrem R 1-03:13:42 1 med0148\nLooking at man squeue, we learn that the default format strings are:
default \"%.18i %.9P %.8j %.8u %.2t %.10M %.6D %R\"\n-l, --long \"%.18i %.9P %.8j %.8u %.8T %.10M %.9l %.6D %R\"\n-s, --steps \"%.15i %.8j %.9P %.8u %.9M %N\"\nThis looks a bit wasteful. Let's cut down on the padding of the job ID and expand on the job name and remove some right paddings.
hpc-login-1:~$ squeue -o \"%.6i %9P %30j %.10u %.2t %.10M %.6D %R %b\" | head\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 3149 medium variant_calling holtgrem_c PD 0:00 1 (Dependency)\n 1177 medium bash jweiner_m R 6-03:35:55 1 med0127\n 1192 medium bash jweiner_m R 5-12:52:11 1 med0127\n 1210 gpu bash hilberta_c R 2-16:14:05 1 med0304\n 1213 long bash schubacm_c R 2-15:25:58 1 med0127\n 2401 gpu bash ramkem_c R 2-10:58:24 1 med0303\n 3063 long bash schubacm_c R 1-09:56:08 1 med0127\n 3066 long bash schubacm_c R 1-09:55:18 1 med0127\n 3147 medium ngs_mapping holtgrem_c R 1-03:16:56 1 med0148\nNow display how many of our internal projects still exist.
hpc-login-1:~$ squeue -o \"%.6i %9P %30j %.10u %.2t %.10M %.6D %10R %s\" | head\nThe next steps are (TODO):
This page describes how to create SLURM job scripts.
SLURM job scripts look as follows. On the top you have lines starting with #SBATCH. These appear as comments to bash scripts. These lines are interpreted by sbatch in the same way as command line arguments. That is, when later submitting the script with sbatch my-job.sh you can either have the parameter to the sbatch call or in the file.
Multi-Node Allocation in Slurm
Classically, jobs on HPC systems are written in a way that they can run on multiple nodes at once, using the network to communicate. Slurm comes from this world and when allocating more than one CPU/core, it might allocate them on different nodes. Please use --nodes=1 to force Slurm to allocate them on a single node.
Creating the Script
host:example$ cat >my-job.sh <<\"EOF\"\n#!/bin/bash\n#\n#SBATCH --job-name=this-is-my-job\n#SBATCH --output=output.txt\n#\n#SBATCH --ntasks=1\n#SBATCH --nodes=1\n#SBATCH --time=10:00\n#SBATCH --mem-per-cpu=100M\n\ndate\n\nhostname\n>&2 echo \"Hello World\"\n\nsleep 1m\n\ndate\nEOF\nAlso see the SLURM Rosetta Stone for more options.
Submit, Look at Queue & Result
host:example$ sbatch script.sh \nSubmitted batch job 315\nhost:example$ squeue -u holtgrem_c\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) \n 315 debug this-is- holtgrem R 0:40 1 med0127 \nhost:example$ sleep 2m\nhost:example$ squeue -u holtgrem_c\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON) \nhost:example$ cat output.txt \nWed Mar 25 13:30:56 CET 2020\nmed0127\nHello World\nWed Mar 25 13:31:56 CET 2020\nMemory allocation is one of the topics that users find confusing most often. This section first gives some technical background and then explains how to implement this properly with Slurm on the BIH HPC.
"},{"location":"slurm/memory-allocation/#technical-background","title":"Technical Background","text":"Technical Background Summary
Main memory used to be one of the most important topics when programming, as computers had so little. There is the infamous quote \"640KB ought ot be enough for anybody\" wrongly attribute to Bill Gates which refers to the fact that early computers could only address that amount of memory. In MS DOS, one had to use special libraries for a program to use more memory. Today, computers are very fast and memory is plentiful and people can (rightfully) forget about memory allocation ... as long as they don't use \"much\" memory by today's standards.
The Linux operating system differentiates between the following types of memory:
Note that above we are talking about processes, not Slurm jobs yet. Let us look at this in detail:
Each program uses some kind of memory management. For example, in C the malloc and free functions manually allocate and free memory while in Java, R, and Python, memory allocation and release is done automatically using a concept called garbage collection. Each program starts with a certain virtual memory size, that is the amount of memory it can address, say 128MB. When the program allocates memory, the memory allocation mechanism will check whether it has sufficient space left. If not, it will request an increase in virtual memory from the operating system, e.g., to 256MB. If this fails then the program can try to handle the error, e.g., terminate gracefully, but many programs will just panic and stop. Otherwise, the program will get access to more memory and happily continue to run.
However, programs can allocate humonguous amounts of virtual memory and only use a little. Memory is organized in \"pages\" (classically these are 4096 bytes each, but can be larger using so-called \"huge page\" features). The operating system tracks which memory pages are actually used by a process. The total size of these pages is called the resident set size: the amount of memory that is actually currently used by a program. Programs can also mark pages as unused again, thus freeing resident memory and can also decrease their virtual memory.
In some cases it is still interesting to use swap memory. Here, the contents of resident memory are copied to disk by the operating system. This process is completely transparent to the program; the data remains available at the original positions in the virtual memory! However, accessing it will take some time as it must be read back into main memory from the disk. In this way, it was possible for a computer with 4MB of RAM and a disk of 100MB to run programs that used 8MB. Of course, this was only really useable for programs that ran in the background. One could really feel the latency if a graphical program was using swapped memory (you could actually hear the hard drive working). Today, swap storage is normally only relevant when put your computer into hibernation. Given the large main memory on the cluster nodes, their small local hard drives (just used for loading the operating system), and the extreme slowness involved in using swapped memory, the BIH HPC nodes have no swap memory allocated.
Most HPC users will also use shared memory, at least implicitly. Whenever a program uses fork to create a subprocess (BTW, this is not a thread), the program can chose to \"copy\" its current address space. The second process then has access to the same memory than the parent process in a copy-on-write fashion. This allows, for example, pre-loading a database, and also allows the use of already loaded library code by the child process as well. If the child process writes to the copy-on-write memory of the parent, the relevant memory page will be copied and attributed to the child.
Two or more processes can share the same memory explicitly. This is usually used for inter-process communication but the Bowtie program uses it for sharing the memory of indices. For example, the Python multiprocessing module will use this, including if you have two MPI processes running on the same host.
Memory is also separated into segments, the most interesting ones are heap and stack memory. For compiled languages, memory can be allocated on either. For C, an int variable will be allocated on the stack. Every time you call a function, a stack frame is created in memory to hold the local variables and other information for the duration of the function execution. The stack thus grows through function calls made by your program and shrinks when the functions return. The stack size for a process is limited (by ulimit -s) and a program that goes too deep (e.g., via infinite recursion) will be terminated by the operating system if it exceeds this limit. Again in C, int * ptr = (int *)malloc(10 * sizeof(int)); will allocate memory for one variable (an integer pointer) on the stack and memory for 10 integers on the heap. When the function returns, the ptr variable on the stack will be freed but to free the array of integers, you'd have to call free(ptr). If the memory is not freed then this constitutes a memory leak, but that is another topic.
Other relevant segments are code, where the compiled code lives, and data, where static data such as strings displayed to the user are stored. As a side node, in interpreted languages such as R or Python, the code and data segments will refer to the code and data of Python while the actual program text will be on the heap.
"},{"location":"slurm/memory-allocation/#interlude-memory-in-java","title":"Interlude: Memory in Java","text":"Memory in Java Summary
-XX:MaxHeapSize=<size> (e.g., <size>=2G) for your program and only tune the other parameters if neededJava's memory management provides for some interesting artifacts. When running simple Java programs, you will never run into this but if you need to use gigabytes of memory in Java then you will have to learn a bit about Java memory management. This is the case when running GATK programs, for example.
As different operating systems handle memory management differently, the Java virtual machine does its own memory management to provide a consistent interface. The following three settings are important in governing memory usage of Java:
-Xmx<size>/-XX:MaxHeapSize=<size> -- the maximal Java heap size-Xms<size>/-XX:InitialHeapSize=<size> -- the initial Java heap size-Xss<size>/-XX:ThreadStackSize=<size> -- maximal stack size available to a Java thread (e.g., the main thread)Above, <size> is a memory specification, either in bytes or with a suffix, e.g., 80M, or 1G.
On startup, Java does roughly the following:
Memory freed by the Java garbage collector can be re-used by other Java objects (rss remains the same) or be freed in the operating system (rss decreases). The Java VM program itself will also consume memory on the OS stack but that is negligible.
Overall, the Java VM needs to store in main memory:
In the BIH HPC context, the following is recommended to:
Memory Allocation in Slurm Summary
--mem=<size> (e.g., <size>=3G) to allocate memory per nodesrun and batch sbatch jobs are governed by Slurm memory allocationOur Slurm configuration uses Linux cgroups to enforce a maximum amount of resident memory. You simply specify it using --mem=<size> in your srun and sbatch command.
In the (rare) case that you provide more flexible number of threads (Slurm tasks) or GPUs, you could also look into --mem-per-cpu and --mem-per-gpu. The official Slurm sbatch manual is quite helpful, as is man sbatch on the cluster command line.
Slurm (or rather Linux via cgroups) will track all memory started by all jobs by your process. If each process works independently (e.g., you put the output through a pipe prog1 | prog2) then the amount of memory consumed will at any given time be the sum of the RSS of both processes at that time. If your program uses fork, which uses memory in a copy-on-write fashion, the shared memory is of course only counted once. Note that Python's multiprocessing does not use copy on write: its data will be explicitly copied and consume additional memory. Refer to the Scipy/Numpy/Pandas etc. documentation on how to achieve parallelism without copying too much data.
The amount of virtual memory that your program can reserve is only \"virtually\" unlimited (pun not intended). However, in practice, the operating system will not like you allocating more than physically available. If your program attempts to allocate more memory than requested via Slurm, your program will be killed.
This is reported to you in the Slurm job output log as something like:
slurmstepd: error: Detected 1 oom-kill event(s) in step <JOB ID>.batch cgroup. Some of your processes may have been killed by the cgroup out-of-memory handler.\nYou can inspect the amount of memory available on each node in total with sinfo --format \"%.10P %.10l %.6D %.6m %N\", as shown below.
$ sinfo --format \"%.10P %.10l %.6D %.6m %N\"\n PARTITION TIMELIMIT NODES MEMORY NODELIST\n debug* 8:00:00 240 128722 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\n medium 7-00:00:00 240 128722 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\n long 28-00:00:0 240 128722 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\n critical 7-00:00:00 176 128722 med[0101-0164,0501-0516,0601-0632,0701-0764]\n highmem 14-00:00:0 4 515762 med[0401-0404]\n gpu 14-00:00:0 4 385215 med[0301-0304]\n mpi 14-00:00:0 240 128722 med[0101-0164,0201-0264,0501-0516,0601-0632,0701-0764]\nMemory Accounting in Slurm Summary
sacct -j JOBID --format=JobID,MaxRSS to display the RSS usage of your programsacct -j JOBID --format=Elapsed,AllocCPUs,TotalCPU to display information about CPU usageWhile Slurm runs your job, it collects information about the job such as the running time, exit status, and memory usage. This information is available through the scheduling system via the squeue and scontrol commands, but only while the job is pending execution, executing, or currently completing. After job completion, the information is only available through the Slurm accounting system.
You can query information about jobs, e.g., using sacct:
$ sacct -j 1607166\n JobID JobName Partition Account AllocCPUS State ExitCode\n------------ ---------- ---------- ---------- ---------- ---------- --------\n1607166 snakejob.+ critical 16 COMPLETED 0:0\n1607166.bat+ batch 16 COMPLETED 0:0\n1607166.ext+ extern 16 COMPLETED 0:0\nThis shows that the job with ID 1607166 with a job ID starting with snakejob. has been run in the critical partition, been allocated 16 cores and had an exit code of 0:0. For technical reasons, there is a batch and an extern sub step. Actually, Slurm makes it possible to run various steps in one batch as documented in the Slurm documentation.
The sacct command has various command-line options that you can read about via man sacct or in the Slurm documentation. We can use --brief/-b to show only a brief summary.
$ sacct -j 1607166 --brief\n JobID State ExitCode\n------------ ---------- --------\n1607166 COMPLETED 0:0\n1607166.bat+ COMPLETED 0:0\n1607166.ext+ COMPLETED 0:0\nSimilarly, you can use --long to display extended information (see the manual for the displayed columns). Very long report lines can be piped into less -S for easier display. You can fine-tune the information to display with a format string to --format:
$ sacct -j 1607166 --format=JobID,ReqMem,MaxRSS,Elapsed,TotalCPU,AllocCPUS\n JobID ReqMem MaxRSS Elapsed TotalCPU AllocCPUS\n------------ --------- ---------- ---------- ---------- ----------\n1607166 60Gn 13:07:31 7-16:21:29 16\n1607166.bat+ 60Gn 4314560K 13:07:31 7-16:21:29 16\n1607166.ext+ 60Gn 0 13:07:31 00:00.001 16\nFrom this command, we can read that we allocate 60GB memory of memory per node (suffix n, here Gn for gigabytes per node) and the maximum RSS is reported as 4.3GB. You can use this information to fine-tune your memory allocations. As a side-remark, a suffic c indicates the memory per core (e.g., that could be60Gc)
Further, the program ran for 13 hours and 7 minutes with allocated 16 CPU cores and consumed a total of 7 days, 16 hours, and 21 minutes of CPU time. Thus, a total of 10,061 CPU minutes were spent in 787 minutes wall-clock time. This yields an overall empirical degree of parallelism of about 10061 / 787 = 14, and a parallel efficiency of 14 / 16 = 88%. The discussion of parallel efficiency is a topic not covered here.
However, you can use the awk script below to compute the empirical parallelism (EmpPar) and the parallel efficiency (ParEff). The script also displays the difference I requested, and used RSS (DiffRSS). The script can be found here.
$ sacct -j 1607166 --format=JobID,ReqMem,MaxRSS,Elapsed,TotalCPU,AllocCPUS \\\n | awk -f quick-sacct.awk\n JobID ReqMem MaxRSS Elapsed TotalCPU AllocCPUS EmpPar ParEff DiffMEM\n------------ ---------- ---------- ---------- ---------- ---------- --------- -------- --------\n1607166 60Gn 13:07:31 7-16:21:29 16 0.00 0.00 -\n1607166.bat+ 60Gn 4314560K 13:07:31 7-16:21:29 16 14.05 0.88 55.89\n1607166.ext+ 60Gn 0 13:07:31 00:00.001 16 0.00 0.00 -\nThe BIH HPC uses the Slurm scheduling system for resource allocation. This section of the manual attempts to give an overview of what scheduling is and how to use the Slurm scheduler. For more detailed information, you will have to refer to the Slurm website and the Slurm man pages (e.g., by entering man sbatch or man srun on the HPC terminal's command line).
For a quick introduction and hands-on examples, please see the manual sections
Also, make sure that you are aware of our How-To: Debug Software and How-To: Debug Software on HPC Systems guides in the case that something goes wrong.
"},{"location":"slurm/overview/#annotated-contents","title":"Annotated Contents","text":"srun -- running parallel jobs nowsbatch -- submission of batch jobsscancel -- stop/kill jobssinfo -- display information about the Slurm clustersqueue -- information about pending and running jbosscontrol -- detailed information (and control)sacct -- access Slurm accounting information (pending, running, and past jobs)Many other facilities run Slurm clusters and make their documentation available on the internet. We list some that we found useful below. However, be aware that Slurm is a highly configurable and extensible system. Other sites may have different configurations and plugins enabled than we have (or might even have written custom plugins that are not available at BIH). In any case, it's always useful to look \"\u00fcber den Tellerrand\".
man Pages - web versions of Unix manual (man) pages.Create an interactive bash session (srun will run bash in real-time, --pty connects its stdout and stderr to your current session).
hpc-login-1:~$ srun --pty bash -i\nmed0740:~$ echo \"Hello World\"\nHello World\nmed0740:~$ exit\nhpc-login-1:~$\nNote you probably want to longer running time for your interactive jobs. This way, your jobs can run for up to 28 days. This will make your job be routed automatically into the long partition as it is the only one that can fit your job.
hpc-login-1:~$ srun --pty --time 28-00 bash -i\nmed0740:~$\nPro-Tip: Using Bash aliases for quick access.
hpc-login-1:~$ alias slogin=\"srun --pty bash -i\"\nhpc-login-1:~$ slogin\nmed0740:~$ exit\nhpc-login-1:~$ cat >>~/.bashrc <<\"EOF\"\n# Useful aliases for logging in via Slurm\nalias slogin=\"srun --pty bash -i\"\nalias slogin-x11=\"srun --pty --x11 bash -i\"\nEOF\nCreate an interactive R session on the cluster (assuming conda is active and the environment my-r is created, e.g., with conda create -n my-r r).
hpc-login-1:~$ conda activate my-r\nhpc-login-1:~$ srun --pty R\nR version 3.6.2 (2019-12-12) -- \"Dark and Stormy Night\"\nCopyright (C) 2019 The R Foundation for Statistical Computing\n[...]\nType 'demo()' for some demos, 'help()' for on-line help, or\n'help.start()' for an HTML browser interface to help.\nType 'q()' to quit R.\n\n\n> Sys.info()[\"nodename\"]\n nodename\n\"med0740\"\n> q()\nSave workspace image? [y/n/c]:\nhpc-login-1:~$\nCreate an interactive iPython session on the cluster (assuming conda is active and the environment my-python is created, e.g., with conda create -n my-python python=3 ipython).
hpc-login-1:~$ conda activate my-python\nhpc-login-1:~$ srun --pty ipython\nPython 3.8.2 | packaged by conda-forge | (default, Mar 5 2020, 17:11:00)\nType 'copyright', 'credits' or 'license' for more information\nIPython 7.13.0 -- An enhanced Interactive Python. Type '?' for help.\n\nIn [1]: import socket; socket.gethostname()\nOut[1]: 'med0740'\n\nIn [2]: exit\nhpc-login-1:~$\nAllocate 4 cores (default is 1 core), and a total of 4GB of RAM on one node (alternatively use --mem-per-cpu to set RAM per CPU); sbatch accepts the same argument.
hpc-login-1:~$ srun --cpus-per-task=4 --nodes=1 --mem=4G --pty bash\nmed0740:~$ export | grep SLURM_CPUS_ON_NODE\n4\nmed0740:~$ your-parallel-script --threads 4\nSubmit an R script to the cluster in batch mode (sbatch schedules the job for later execution).
hpc-login-1:~$ cat >job-script.sh <<\"EOF\"\n#!/bin/bash\necho \"Hello, I'm running on $(hostname) and it's $(date)\"\nEOF\nhpc-login-1:~$ sbatch job-script.sh\nSubmitted batch job 7\n\n# Some time later:\nhpc-login-1:~$ cat slurm-7.out\nHello, I'm running on med0740 and it's Fri Mar 6 07:36:42 CET 2020\nhpc-login-1:~$\nHint
Read this in particular if you want to know why your job does not get scheduled and you see Reason=ReqNodeNotAvail,_Reserved_for_maintenance in scontrol show job .
Administration registers maintenances with the Slurm scheduler as so-called reservations. You can see the current reservations with scontrol show reservation. The following is a scheduled reservation affecting ALL nodes of the cluster.
# scontrol show reservation\nReservationName=root_13 StartTime=2021-09-07T00:00:00 EndTime=2021-09-09T00:00:00 Duration=2-00:00:00\n Nodes=hpc-cpu-[1-36],med[0101-0116,0201-0264,0301-0304,0401-0404,0501-0516,0601-0632,0701-0764]\n NodeCnt=236 CoreCnt=5344 Features=(null) PartitionName=(null)\n Flags=MAINT,IGNORE_JOBS,SPEC_NODES,ALL_NODES TRES=cpu=10176\n Users=root Groups=(null) Accounts=(null) Licenses=(null) State=INACTIVE BurstBuffer=(null) Watts=n/a\n MaxStartDelay=(null)\nYou will also be notified when logging into the login nodes, e.g.,
--\n ***NOTE: 1 scheduled maintenance(s)***\n\n 1: 2021-09-07 00:00:00 to 2021-09-09 00:00:00 ALL nodes\n\nYou jobs do not start because of \"Reserved_for_maintenance\"?\nSlurm jobs will only start if they do not overlap with scheduled reservations.\nMore information:\n\n - https://bihealth.github.io/bih-cluster/slurm/reservations/\n - https://bihealth.github.io/bih-cluster/admin/maintenance/\n--\nMaintenance reservations will block the affected nodes (or even the whole cluster) for jobs. If there is a maintenance in one week then your job must have an end time before the reservation starts. By this, the job gives a guarantee to the scheduler that it will not interfer with the maintenance reservation.
For example, scontrol show job JOBID might report the following
JobId=4011580 JobName=snakejob\n UserId=USER(UID) GroupId=GROUP(GID) MCS_label=N/A\n Priority=1722 Nice=0 Account=GROUP QOS=normal\n JobState=PENDING Reason=ReqNodeNotAvail,_Reserved_for_maintenance Dependency=(null)\n Requeue=1 Restarts=0 BatchFlag=1 Reboot=0 ExitCode=0:0\n RunTime=00:00:00 TimeLimit=28-00:00:00 TimeMin=N/A\n SubmitTime=2021-08-30T09:01:01 EligibleTime=2021-08-30T09:01:01\n AccrueTime=2021-08-30T09:01:01\n StartTime=2021-09-09T00:00:00 EndTime=2021-10-07T00:00:00 Deadline=N/A\n SuspendTime=None SecsPreSuspend=0 LastSchedEval=2021-08-30T10:20:40\n Partition=long AllocNode:Sid=172.16.35.153:5453\n ReqNodeList=(null) ExcNodeList=(null)\n NodeList=(null)\n NumNodes=1-1 NumCPUs=8 NumTasks=8 CPUs/Task=1 ReqB:S:C:T=0:0:*:*\n TRES=cpu=8,mem=4G,node=1,billing=8\n Socks/Node=* NtasksPerN:B:S:C=0:0:*:* CoreSpec=*\n MinCPUsNode=1 MinMemoryNode=4G MinTmpDiskNode=0\n Features=(null) DelayBoot=00:00:00\n OverSubscribe=OK Contiguous=0 Licenses=(null) Network=(null)\n Power=\n NtasksPerTRES:0\nLook out for the Reason line:
Reason=ReqNodeNotAvail,_Reserved_for_maintenance\nThis job is scheduled to run up to 4 weeks and has been submitted on 2021-08-30.
Right now the following reservation is active
# scontrol show reservation\nReservationName=root_13 StartTime=2021-09-07T00:00:00 EndTime=2021-09-09T00:00:00 Duration=2-00:00:00\n Nodes=hpc-cpu-[1-36],med[0101-0116,0201-0264,0301-0304,0401-0404,0501-0516,0601-0632,0701-0764]\n NodeCnt=236 CoreCnt=5344 Features=(null) PartitionName=(null)\n Flags=MAINT,IGNORE_JOBS,SPEC_NODES,ALL_NODES TRES=cpu=10176\n Users=root Groups=(null) Accounts=(null) Licenses=(null) State=INACTIVE BurstBuffer=(null) Watts=n/a\n MaxStartDelay=(null)\nThus, the scheduler decided to set a StartTime of the job to 2021-09-09T00:00:00, which is the end time of the reservation. Effectively, the job is forced to run outside the maintenance reservation.
You can resolve this by using a --time= parameter to srun or sbatch such that the job ends before the maintenance reservation starts.
Rosetta Stone?
The Rosetta Stone is a stone slab that carries the same text in Egyptian hieroglyphs and ancient Greek. This was key for decyphering Egyptian hieroglyphs in the 18th century. Nowadays, the term is often used to label translation tables such as the one below.
The table below shows some SGE commands and their Slurm equivalents.
User Command SGE Slurm remote loginqrsh/qlogin srun --pty bash run interactively N/A srun --pty program submit job qsub script.sh sbatch script.sh delete job qdel job-id scancel job-id job status by job id N/A squeue --job job-id detailed job status qstat -u '*' -j job-id sstat job-id job status of your jobs qstat squeue --me job status by user qstat -u user squeue -u user hold job qhold job-id scontrol hold job-id release job qrls job-id scontrol release job-id queue list qconf -sql scontrol show partitions node list qhost sinfo -N OR scontrol show nodes cluster status qhost -q sinfo show node resources N/A sinfo \"%n %G\" Job Specification SGE Slurm script directive marker #$ #SBATCH (run in queue) -q queue -p queue allocated nodes N/A -N min[-max] allocate cores -pe smp count -n count limit running time -l h_rt=time -t days-hh:mm:s redirectd stdout -o file -o file redirect stderr -e file -e file combine stdout/stderr -j yes -o without -e copy environment -V --export=ALL\\|NONE\\|variables email notification -m abe --mail-type=events send email to -M email --mail-user=email job name -N name --job-name=name restart job -r yes|no --requeue|--no-requeue working directory -wd path --workdir run exclusively -l exclusive --exclusive OR --shared allocate memory -l h_vmem=size --mem=mem OR --mem-per-cpu=mem wait for job -hold_jid jid --depend state:job select target host -l hostname=host1\\|host1 --nodelist=nodes AND/OR --exclude allocate GPU -l gpu=1 --gres=gpu:tesla:count or --gres=gpu:a40:count"},{"location":"slurm/snakemake/","title":"Snakemake with Slurm","text":"This page describes how to use Snakemake with Slurm.
"},{"location":"slurm/snakemake/#prerequisites","title":"Prerequisites","text":"source miniforge/bin/activate.We first create a new environment snakemake-slurm and activate it. We need the snakemake package for this. For snakemake 8, we additionally need snakemake-executor-plugin-slurm.
host:~$ conda create -y -n snakemake-slurm 'snakemake>=8.24.1' snakemake-executor-plugin-slurm\n[...]\n#\n# To activate this environment, use\n#\n# $ conda activate snakemake-slurm\n#\n# To deactivate an active environment, use\n#\n# $ conda deactivate\nhost:~$ conda activate snakemake-slurm\n(snakemake-slurm) host:~$\nWe create a workflow and ensure that it works properly with multi-threaded Snakemake (no cluster submission here!)
host:~$ mkdir -p snake-slurm\nhost:~$ cd snake-slurm\nhost:snake-slurm$ cat >Snakefile <<\"EOF\"\nrule default:\n input: \"the-result.txt\"\n\nrule mkresult:\n output: \"the-result.txt\"\n shell: r\"sleep 2m; touch the-result.txt\"\nEOF\nhost:snake-slurm$ snakemake --cores=1\n[...]\nhost:snake-slurm$ ls\nSnakefile the-result.txt\nhost:snake-slurm$ rm the-result.txt\nSimply use snakemake --profile=cubi-v1 and the Snakemake resource configuration as shown below. This works for both snakemake version 7 and 8.
(For version 7, you can also use snakemake --cluster='sbatch ...' command instead, but this is discouraged.)
Note that we sneaked in a sleep 2m? In a second terminal session, we can see that the job has been submitted to SLURM indeed.
host:~$ squeue -u holtgrem_c\n JOBID PARTITION NAME USER ST TIME NODES NODELIST(REASON)\n 325 debug snakejob holtgrem R 0:47 1 med0127\nThe cubi-v1 profile (stored in /etc/xdg/snakemake/cubi-v1 on all cluster nodes) supports the following specification in your Snakemake rule:
threads: the number of threads to execute the job onresources.mem/resources.mem_mb: the memory to allocate for the whole job, OR resources.mem_per_thread: the memory to allocate for each thread.resources.time: the running time of the rule, in a syntax supported by Slurm, e.g. HH:MM:SS or D-HH:MM:SSresources.partition: the partition to submit your job into (Slurm will pick a fitting partition for you by default)resources.nodes: the number of nodes to schedule your job on (defaults to 1 and you will want to keep that value unless you want to use MPI)You will need Snakemake >=7.0.2 for this.
For more information on resources, see snakemake resources and slurm resources.
Here is how to call Snakemake:
# snakemake --profile=cubi-v1 --jobs 1\n--sdm conda (snakemake 8) or --use-conda (snakemake 7). To set rule-specific resources:
rule myrule:\n threads: 1\n resources:\n mem='8G',\n time='04:00:00',\n input: # ...\n output: # ...\n shell: # ...\nYou can combine this with Snakemake resource callables, of course:
def myrule_mem(wildcards, attempt):\n mem = 2 * attempt\n return '%dG' % mem\n\nrule snps:\n threads: 1\n resources:\n mem=myrule_mem,\n time='04:00:00',\n input: # ...\n output: # ...\n shell: # ...\nBy default, slurm will write log files into the working directory of snakemake, which will look like slurm-$jobid.out.
To change this behaviour, the environment variable SBATCH_DEFAULTS can be set to re-route the --output parameter. If you want to write your files into slurm_logs with a filename pattern of $name-$jobid for instance, consider the following snippet for your submission script:
#!/bin/bash\n#\n#SBATCH --job-name=snakemake_main_job\n#SBATCH --ntasks=1\n#SBATCH --nodes=1\n#SBATCH --time=48:10:00\n#SBATCH --mem-per-cpu=300M\n#SBATCH --output=slurm_logs/%x-%j.log\n\nmkdir -p slurm_logs\nexport SBATCH_DEFAULTS=\" --output=slurm_logs/%x-%j.log\"\n\ndate\nsrun snakemake --use-conda -j1 --profile=cubi-v1\ndate\nThe name of the snakemake slurm job will be snakemake_main_job, the name of the jobs spawned from it will be called after the rule name in the Snakefile.
This section describes how Slurm handles temporary files on the local disk.
Temporary Files Best Practices
See Best Practices: Temporary Files for information how to use temporary files effectively.
"},{"location":"slurm/temporary-files/#slurm-behaviour","title":"Slurm Behaviour","text":"Our Slurm configuration has the following behaviour.
"},{"location":"slurm/temporary-files/#environment-variable-tmpdir","title":"Environment Variable TMPDIR","text":"Slurm itself will by default not change the TMPDIR environment variable but retain the variable's value from the srun or sbatch call.
/tmp Directories","text":"The only place where users can write data to on local storage of the compute nodes is /tmp.
Storage is a consumable shared resource as the storage used by one job cannot use another job. It is thus critical that Slurm cleans up after each job such that all space on the local node is available to the next job. This is done using the job_container/tmpfs Slurm plugin.
This plugin creates a so-called Linux namespace for each job and creates a bind mount of /tmp to a location on the local storage. This mount is only visible to the currently running job and each job, even of the same user, get their own /tmp. After a job terminates, Slurm will remove the directory and all of its content.
There is a notable exception. If you use ssh to connect to a node rather than using srun or sbatch, you will see the system /tmp directory and can also write to it. This usage of storage is not tracked and consequently you can circumvent the Slurm quota management. Using /tmp in this fashion (i.e., outside of Slurm-controlled jobs) is prohibited. If it cannot be helped (e.g., if you need to run some debugging application that needs to create FIFO or socket files) then keep usage of /tmp outside of Slurm job below 100MB.
localtmp","text":"Enforcing localtmp Gres
From January 31, we will enforce the allocated storage in /tmp on the local disk with quotas. Jobs writing to /tmp beyond the quota in the job allocation will not function properly and probably crash with \"out of disk quota\" messages.
Slurm tracks the available local storage above 100MB on nodes in the localtmp generic resource (aka Gres). The resource is counted in steps of 1MB, such that a node with 350GB of local storage would look as follows in scontrol show node:
hpc-login-1 # scontrol show node hpc-cpu-1\nNodeName=hpc-cpu-1 Arch=x86_64 CoresPerSocket=24\n [...]\n Gres=localtmp:350K\n [...]\n CfgTRES=cpu=96,mem=360000M,billing=96,gres/localtmp=358400\n [...]\nEach job is automaticaly granted 100MB of storage on the local disk which is sufficient for most standard programs. If your job needs more temporary storage then you should either
$HOME/scratch volume (see Best Practices: Temporary Files)localtmp generic resource (described here)You can allocate the resource with --gres=localtmp:SIZE where SIZE is given in MB.
hpc-login-1 # srun --gres=localtmp:100k --pty bash -i\nhpc-cpu-1 # scontrol show node hpc-cpu-1\nNodeName=hpc-cpu-1 Arch=x86_64 CoresPerSocket=24\n [...]\n Gres=localtmp:250K\n [...]\n CfgTRES=cpu=96,mem=360000M,billing=96,gres/localtmp=358400\n [...]\n AllocTRES=cpu=92,mem=351G,gres/localtmp=102400\n [...]\nThe first output tells us about the resource configured to be available to user jobs and the last line show us that 100k=102400 MB of local storage are allocated.
You can also see the used resources in the details of your job:
scontrol show job 14848\nJobId=14848 JobName=example.sh\n [...]\n TresPerNode=gres:localtmp:100k\nMake sure to connect to the login node with X11 forwarding.
host:~$ ssh -X -l user_c hpc-login-1.cubi.bihealth.org\nOnce connected to the login node, pass the --x11 flag.
hpc-login-1:~$ srun --pty --x11 xterm\nWe set quite restrictive quotas for user homes, but in exchange you get file system snapshots and mirroring. Your home folder should therefore only be used for scripts, your user config, and other small files. Everything else should be stored in the work or scratch subdirectories, which effectively link to your group's shared storage space. This document describes some common pitfalls and how to circumvent them.
Hint
The tilde character (~) is shorthand for your home directory.
Various programs are used to depositing large folders in a user's home and can quickly use up your allotted storage quota. These include:
~/.local/lib/python*~/R/x86_64-pc-linux-gnu-library~/ondemandPlease note that directories whose name is starting with a dot are not shown by the normal ls command, but require the ls -a flag. You can search your home folder for large directories like so:
$ du -shc ~/.* ~/* --exclude=.. --exclude=.\nYou should move these locations to your work folder and create symbolic links in their place. Conda installations should be installed in work from the very beginning as they do not react well to being moved around.
Here is an example for the .local folder.
$ mv ~/.local ~/work/.local\n$ ln -s ~/work/.local ~/.local\nAnother usual culprit is the hidden .cache directory which contains temporary files. This folder can be moved to the scratch volume in a similar manner as described above.
$ mv ~/.cache ~/scratch/.cache\n$ ln -s ~/scratch/.cache ~/.cache\nImportant
Files placed in your scratch directory will be automatically removed after 2 weeks. Do not place any valuable files in there.
Please use hpc-transfer-1 and hpc-transfer-2 for moving large amounts of files. This not only leaves the compute notes available for actual computation, but also has no risk of your jobs being killed by Slurm. You should also use tmux to not risk connection loss during long running transfers.
Define source and target location and copy contents. Please replace the parts in curly brackets with your actual folder names. It is important to end paths with a trailing slash (/) as this is interpreted by sync as \u201call files in this folder\u201d.
$ SOURCE=/data/gpfs-1/work/projects/{my_project}/\n$ TARGET=/data/cephfs-2/unmirrored/projects/{my-project}/\n$ rsync -ahP --stats --dry-run $SOURCE $TARGET\nRemove the --dry-run flag to start the actual copying process.
Important
File ownership information will be lost during this process. This is due to non-root users not being allowed to change ownership of arbitrary files. If this is a problem for you, please contact our admins again after completing this step.
Perform a second rsync to check if all files were successfully transferred. Paranoid users might want to add the --checksum flag to rsync or use hashdeep. Please note the flag --remove-source-files which will do exactly as the name suggests, but leaves empty directories behind.
$ rsync -ahX --stats --remove-source-files --dry-run $SOURCE $TARGET\n--dry-run flag to start the actual deletion.$ find $SOURCE -type f | wc -l\n0\n$ rm -r $SOURCE\nWarning
When defining your SOURCE location, do not use the * wildcard character. It will not match hidden (dot) files and leave them behind. Its better to use a trailing slash which matches \u201cAll files in this folder\u201d.
All files within your own work directory can be transferred as follows. Please replace parts in curly braces with your cluster user name.
$ SOURCE=/data/gpfs-1/work/users/{username}/\n$ TARGET=/data/cephfs-1/home/users/{username}/work/\n$ rsync -ahP --stats --dry-run $SOURCE $TARGET\nNote
The --dry-run flag lets you check that rsync is working as expected without copying any files. Remove it to start the actual transfer.
Perform a second rsync to check if all files were successfully transferred. Paranoid users might want to add the --checksums flag or use hashdeep. Please note the flag --remove-source-files which will do exactly as the name suggests, but leaves empty directories behind.
$ rsync -ahP --stats --remove-source-files --dry-run $SOURCE $TARGET\n$ find $SOURCE -type f | wc -l\n0\nConda installations tend not to react well to moving their main folder from its original location. There are numerous ways around this problem which are described here.
A simple solution we can recommend is this:
Install a fresh version of conda or mamba in your new work folder. Don't forget to first remove the conda init block in ~/.bashrc.
$ nano ~/.bashrc\n$ conda init\n$ conda config --set auto_activate_base false\nYou can then use your new conda to export your old environments by specifying a full path like so:
$ conda env export -p /fast/work/user/$USER/miniconda/envs/<env_name> -f <env_name>.yaml\n--no-builds flag. Finally re-create your old environments from the yaml files:
$ conda env create -f {environment.yml}\nOutdated
This document is only valid for the old, third-generation file system and will be removed soon. Quotas of our new CephFS storage are communicated via the HPC Access web portal.
As described elsewhere, all data in your user, group, and project volumes is subject to quotas. This page quickly shows how to query for the current usage of data volume and file counts for your user, group, and projects.
"},{"location":"storage/querying-storage/#query-for-user-data-and-file-usage","title":"Query for User Data and File Usage","text":"The file /etc/bashrc.gpfs-quota contains some Bash functions that you can use for querying the quota usage. This file is automatically sourced in all of your Bash sessions.
For querying your user's data and file usage, enter the following command:
# bih-gpfs-quota-user holtgrem_c\nYou will get a report as follows. As soon as usage reaches 90%, the data/file usage will be highlighted in yellow. If you pass 99%, the data/file usage will be highlighted in red.
=================================\nQuota Report for: user holtgrem_c\n=================================\n\n DATA quota GR- FILES quota GR-\nENTITY NAME FSET USED SOFT HARD ACE USED SOFT HARD ACE\n------- ---------- ------- ----- ---- ----- ----- --- ----- ---- ----- ----- ---\nusers holtgrem_c home 103M 10% 1.0G 1.5G - 2.5k 25% 10k 12k -\nusers holtgrem_c work 639G 62% 1.0T 1.1T - 1.0M 52% 2.0M 2.2M -\nusers holtgrem_c scratch 42G 0% 200T 220T - 207k 0.1% 200M 220M -\n[...]\n# bih-gpfs-report-quota group ag_someag\n=================================\nQuota Report for: group ag_someag\n=================================\n\n DATA quota GR- FILES quota GR-\nENTITY NAME FSET USED SOFT HARD ACE USED SOFT HARD ACE\n------- ---------- ------- ----- ---- ----- ----- --- ----- ---- ----- ----- ---\ngroups ag_someag home 0 0% 1.0G 1.5G - 4 0% 10k 12k -\ngroups ag_someag work 349G 34% 1.0T 1.5T - 302 0% 2.0M 2.2M -\ngroups ag_someag scratch 0 0% 200T 220T - 1 0% 200M 220M -\n\n[...]\n# bih-gpfs-report-quota project someproj\n==================================\nQuota Report for: project someproj\n==================================\n\n DATA quota GR- FILES quota GR-\nENTITY NAME FSET USED SOFT HARD ACE USED SOFT HARD ACE\n------- ---------- ------- ----- ---- ----- ----- --- ----- ---- ----- ----- ---\ngroups someproj home 0 0% 1.0G 1.5G - 4 0% 10k 12k -\ngroups someproj work 349G 34% 1.0T 1.5T - 302 0% 2.0M 2.2M -\ngroups someproj scratch 0 0% 200T 220T - 1 0% 200M 220M -\n\n[...]\nThe scratch space is automatically cleaned up nightly with the following mechanism.
scratch folder are created and retained for 3 days.Warning
We specifically use the mtime attribute to determine if files in scratch should be cleaned up. Copying or downloading files to scratch while preserving the original mtime might lead to unexpected results.
This document describes the forth iteration of the file system structure on the BIH HPC cluster. It was made necessary because the previous file system was no longer supported by the manufacturer and we since switched to distributed Ceph storage.
Important
For now, the old, third-generation file system is still mounted at /fast. It will be decommissioned soon, please consult this document describing the migration process!
There are the following three entities on the cluster:
Each user, group, and project can have storage folders in different locations.
"},{"location":"storage/storage-locations/#data-types-and-storage-tiers","title":"Data Types and Storage Tiers","text":"Files stored on the HPC fall into one of three categories:
Home folders store programs, scripts, and user config i.\u00a0e. long-lived and very important files. Loss of this data requires to redo manual work (like programming).
Work folders store data of potentially large size which has a medium life time and is important. Examples are raw sequencing data and intermediate results that are to be kept (e.\u00a0g. sorted and indexed BAM files). Work data requires time-consuming actions to be restored, such as downloading large amounts of data or long-running computation.
Scratch folder store temporary files with a short life-time. Examples are temporary files (e.\u00a0g. unsorted BAM files). Scratch data is created to be removed eventually.
Ceph storage comes in two types which differ in their I/O speed, total capacity, and cost. They are called Tier 1 and Tier 2 and sometimes hot storage and warm storage. In the HPC filesystem they are mounted in /data/cephfs-1 and /data/cephfs-2.
Storage quotas are imposed in these locations to restrict the maximum size of folders. Amount and utilization of quotas is communicated via the HPC Access web portal.
"},{"location":"storage/storage-locations/#home-directories","title":"Home Directories","text":"Location: /data/cephfs-1/home/
Only users have home directories on Tier 1 storage. This is the starting point when starting a new shell or SSH session. Important config files are stored here as well as analysis scripts and small user files. Home folders have a strict storage quota of 1\u00a0GB.
"},{"location":"storage/storage-locations/#work-directories","title":"Work Directories","text":"Location: /data/cephfs-1/work/
Groups and projects have work directories on Tier 1 storage. User home folders contain a symlink to their respective group's work folder. Files shared within a group/project are stored here as long as they are in active use. Work folders are generally limited to 1\u00a0TB per group. Project work folders are allocated on an individual basis.
"},{"location":"storage/storage-locations/#scratch-space","title":"Scratch Space","text":"Location: /data/cephfs-1/scratch/
Groups and projects have scratch space on Tier 1 storage. User home folders contain a symlink to their respective group's scratch space. Meant for temporary, potentially large data e.\u00a0g. intermediate unsorted or unmasked BAM files, data downloaded from the internet etc. Scratch space is generally limited to 10\u00a0TB per group. Projects are allocated scratch on an individual basis. Files in scratch will be automatically removed 2 weeks after their creation.
"},{"location":"storage/storage-locations/#tier-2-storage","title":"Tier 2 Storage","text":"Location: /data/cephfs-2/
This is where big files go when they are not in active use. Groups are allocated 10 TB of Tier 2 storage by default. File quotas here can be significantly larger as space is much cheaper and more abundant than on Tier 1.
Note
Tier 2 storage is currently not accessible from HPC login nodes.
"},{"location":"storage/storage-locations/#overview","title":"Overview","text":"Tier Function Path Default Quota 1 User home/data/cephfs-1/home/users/<user> 1 GB 1 Group work /data/cephfs-1/work/groups/<group> 1 TB 1 Group scratch /data/cephfs-1/scratch/groups/<group> 10 TB 1 Project work /data/cephfs-1/work/projects/<project> On request 1 Project scratch /data/cephfs-1/scratch/projects/<project> On request 2 Group /data/cephfs-2/unmirrored/groups/<group> 10 TB 2 Project /data/cephfs-2/unmirrored/projects/<project> On request 2 Group /data/cephfs-2/mirrored/groups/<group> On request 2 Project /data/cephfs-2/mirrored/projects/<project> On request"},{"location":"storage/storage-locations/#snapshots-and-mirroring","title":"Snapshots and Mirroring","text":"Snapshots are incremental copies of the state of the data at a particular point in time. They provide safety against various \"Ops, did I just delete that?\" scenarios, meaning they can be used to recover lost or damaged files. Depending on the location and Tier, CephFS creates snapshots in different frequencies and retention plans.
Location Path Retention policy Mirrored User homes/data/cephfs-1/home/users/ Hourly for 48 h, daily for 14 d yes Group/project work /data/cephfs-1/work/ Four times a day, daily for 5 d no Group/project scratch /data/cephfs-1/scratch/ Daily for 3 d no Group/project mirrored /data/cephfs-2/mirrored/ Daily for 30 d, weekly for 16 w yes Group/project unmirrored /data/cephfs-2/unmirrored/ Daily for 30 d, weekly for 16 w no Some parts of Tier 1 and Tier 2 snapshots are also mirrored into a separate fire compartment within the data center. This provides an additional layer of security i.\u00a0e. physical damage to the servers.
"},{"location":"storage/storage-locations/#accessing-snapshots","title":"Accessing Snapshots","text":"To access snapshots simply navigate to the .snap/ sub-folder of the respective location. This special folder exists on all levels of the CephFS file hierarchy, so even in your user home directory. Inside you will find one folder per snapshot created and in those a complete replica of the respective folder at the time of snapshot creation.
For example:
/data/cephfs-1/home/.snap/<some_snapshot>/users/<your_user>/ same as:/data/cephfs-1/home/users/<your_user>/.snap/<some_snapshot>/data/cephfs-1/work/.snap/<some_snapshot>/groups/<your_group>//data/cephfs-2/unmirrored/.snap/<some_snapshot>/projects/<your_project>/Here is a simple example of how to restore a file:
$ cd /data/cephfs-2/unmirrored/groups/cubi/.snap/scheduled-2024-03-11-00_00_00_UTC/\n$ ls -l\nimportant_file.txt\n$ cp important_file.txt /data/cephfs-2/unmirrored/groups/cubi/\n/data/cephfs-1Important
We will remove access to /fast on most cluster nodes following September 30th.
Files on the cluster's main storage /data/gpfs-1 aka. /fast will move to a new file system. That includes users' home directories, work directories, and work-group directories. Once files have been moved to their new locations, /fast will be retired.
Simultaneously we will move towards a more unified naming scheme for project and group folder names. From now on, all such folders names shall be in kebab-case. This is Berlin after all. Group folders will also be renamed, removing the \"ag_\" prefix.
Detailed communication about the move will be communicated via the cluster mailinglist and the user forum. For technical help, please consult the Data Migration Tips and tricks.
"},{"location":"storage/storage-migration/#why-is-this-happening","title":"Why is this happening?","text":"/fast is based on a high performance proprietary hardware (DDN) & file system (GPFS). The company selling it has terminated support which also means buying replacement parts will become increasingly difficult.
There are two file systems set up to replace /fast, named Tier 1 and Tier 2 after their difference in I/O speed:
/fast ever was, but it only has about 75\u00a0% of its usable capacity.The Hot storage Tier 1 is reserved for files requiring frequent random access, user homes, and scratch. Tier 2 (Warm storage) should be used for everything else. Both file systems are based on the open-source, software-defined Ceph storage platform and differ in the type of drives used. Tier 1 or Cephfs-1 uses NVME SSDs and is optimized for performance, Tier 2 or Cephfs-2 used traditional hard drives and is optimized for cost.
So these are the three terminologies in use right now:
/data/cephfs-1/data/cephfs-2More information about CephFS can be found here.
"},{"location":"storage/storage-migration/#new-file-locations","title":"New file locations","text":"Naturally, paths are going to change after files move to their new location. Due to the increase in storage quality options, there will be some more folders to consider.
"},{"location":"storage/storage-migration/#users","title":"Users","text":"/data/cephfs-1/home/users/<user>/data/cephfs-1/work/groups/<doe>/users/<user>/data/cephfs-1/scratch/groups/<doe>/users/<user>Important
User work & scratch spaces are now part of the user's group folder. This means, groups need to coordinate internally to distribute their allotted quota according to each user's needs.
The implementation is done via symlinks created by default when the user account is moved to its new destination:
~/work -> /data/cephfs-1/work/groups/<group>/users/<user>~/scratch -> /data/cephfs-1/scratch/groups/<group>/users/<user>/data/cephfs-1/work/groups/<group>/data/cephfs-1/scratch/groups/<group>/data/cephfs-2/unmirrored/groups/<group>/data/cephfs-1/work/projects/<project>/data/cephfs-1/scratch/projects/<project>Space on Tier 1 is limited. Your colleagues, other cluster users, and admins will be very grateful if you use it only for files you actively need to perform read/write operations on. This means main project storage should probably always be on Tier 2 with workflows to stage subsets of data onto Tier 1 for analysis.
These examples are based on our experience of processing diverse NGS datasets. Your mileage may vary but there is a basic principle that remains true for all projects.
"},{"location":"storage/storage-migration/#dna-sequencing-wes-wgs","title":"DNA sequencing (WES, WGS)","text":"Typical Whole Genome Sequencing data of a human sample at 100x coverage requires about 150 GB of storage, Whole Exome Sequencing files occupy between 6 and 30 GB. These large files require considerable I/O resources for processing, in particular for the mapping step. A prudent workflow for these kind of analysis would therefore be the following:
fastqs) from the Tier 2 location to Tier 1. seqtk is your friend!fastq files from Tier 2 to Tier 1. Run the your scripts on the whole dataset, and copy the results (bam or cram files) back to Tier 2.Tip
Don't forget to use your scratch area for transient operations, for example to sort your bam file after mapping. More information on how to efficiently set up your temporary directory here.
Analysis of RNA expression datasets are typically a long and iterative process, where the data must remain accessible for a significant period. However, there is usually no need to keep raw data files and mapping results available once the gene & transcripts counts have been generated. The count files are much smaller than the raw data or the mapped data, so they can live longer on Tier 1.
A typical workflow would be:
fastq files from Tier 2 to Tier 1.salmon or STAR, and store the results on Tier 2.R, using tximport and DESeq2 or featureCounts & edgeR, for example.R objects) and the output of salmon, STAR, or any mapper/aligner of your choice to Tier 2.Tip
If using STAR, don't forget to use your scratch area for transient operations. More information on how to efficiently set up your temporary directory here
The analysis workflow of bulk RNA & single cell dataset is conceptually similar: Large raw files need to be processed once and only the outcome of the processing (gene counts matrices) are required for downstream analysis. Therefore, a typical workflow would be:
fastq files from Tier 2 to Tier 1.Cell Ranger or alevin-fry, perform count matrix QC and store the results on Tier 2.seurat, scanpy, or Loupe Browser.There is no obvious workflow that covers most used cases for machine learning. However,
/fast to CephFS","text":"Best practices and tools will be provided.
"}]} \ No newline at end of file diff --git a/sitemap.xml b/sitemap.xml index 4fd5bbd5..b37c9005 100644 --- a/sitemap.xml +++ b/sitemap.xml @@ -2,462 +2,462 @@scancel
- December 5, 2024
+ December 9, 2024
diff --git a/slurm/commands-scontrol/index.html b/slurm/commands-scontrol/index.html
index d93cdf6e..e7938534 100644
--- a/slurm/commands-scontrol/index.html
+++ b/slurm/commands-scontrol/index.html
@@ -3243,7 +3243,7 @@ localtmp
- December 5, 2024
+ December 9, 2024
diff --git a/slurm/x11/index.html b/slurm/x11/index.html
index 7aeb3968..928440d1 100644
--- a/slurm/x11/index.html
+++ b/slurm/x11/index.html
@@ -3125,7 +3125,7 @@ Slurm and X11
- December 5, 2024
+ December 9, 2024
diff --git a/storage/home-quota/index.html b/storage/home-quota/index.html
index 369e7027..9f343ce2 100644
--- a/storage/home-quota/index.html
+++ b/storage/home-quota/index.html
@@ -3230,7 +3230,7 @@ Temporary Files
- December 5, 2024
+ December 9, 2024
diff --git a/storage/migration-faq/index.html b/storage/migration-faq/index.html
index 4093b141..a8e65671 100644
--- a/storage/migration-faq/index.html
+++ b/storage/migration-faq/index.html
@@ -3341,7 +3341,7 @@ Conda environments
- December 5, 2024
+ December 9, 2024
diff --git a/storage/querying-storage/index.html b/storage/querying-storage/index.html
index 69578c5a..99bf208e 100644
--- a/storage/querying-storage/index.html
+++ b/storage/querying-storage/index.html
@@ -3269,7 +3269,7 @@ Query for Project Data and File U
- December 5, 2024
+ December 9, 2024
diff --git a/storage/scratch-cleanup/index.html b/storage/scratch-cleanup/index.html
index cd719a5f..306adb2a 100644
--- a/storage/scratch-cleanup/index.html
+++ b/storage/scratch-cleanup/index.html
@@ -3131,7 +3131,7 @@ Automated Cleanup of Scratch
- December 5, 2024
+ December 9, 2024
diff --git a/storage/storage-locations/index.html b/storage/storage-locations/index.html
index 1fbbb4ef..4efc6a0d 100644
--- a/storage/storage-locations/index.html
+++ b/storage/storage-locations/index.html
@@ -3643,7 +3643,7 @@ Tier 2
- December 5, 2024
+ December 9, 2024
diff --git a/storage/storage-migration/index.html b/storage/storage-migration/index.html
index d4aac00e..b4f52dfc 100644
--- a/storage/storage-migration/index.html
+++ b/storage/storage-migration/index.html
@@ -3682,7 +3682,7 @@ Data migration process f
- December 5, 2024
+ December 9, 2024