Issue96 #110
Issue96 #110
Conversation
| { | ||
| fprintf(stderr, "[%s:%d][ERROR] unable to insert send/recv counts\n", __FILE__, __LINE__); | ||
| MPI_Abort(MPI_COMM_WORLD, 1); | ||
| } | ||
| #endif // ((ENABLE_RAW_DATA || ENABLE_PER_RANK_STATS || ENABLE_VALIDATION) && ENABLE_COMPACT_FORMAT) | ||
|
|
||
| #if ((ENABLE_RAW_DATA || ENABLE_PER_RANK_STATS || ENABLE_VALIDATION) && !ENABLE_COMPACT_FORMAT) | ||
| save_counts(sbuf, rbuf, sizeof(sendtype), sizeof(recvtype), comm_size, avCalls); | ||
| int s_dt_size, r_dt_size; | ||
| MPI_Type_size(sendtype, &s_dt_size); |
There was a problem hiding this comment.
This is the same as for alltoall.c, which may of course be what you want. For alltoall it is not necesary to gather the send types because the counts are gathered and bocksize = sendtypesize * sendcount is a constant, so having gathered the sendcounts the sendtypesize used at each node can be inferred. Does something similar apply to the more complicated alltollv - I have not worked it out yet but man for alltoallv says this, which may be helpful:
"When a pair of processes exchanges data, each may pass different element count and datatype arguments so long as the sender specifies the same amount of data to send (in bytes) as the receiver expects to receive.
"Note that process i may send a different amount of data to process j than it receives from process j. Also, a process may send entirely different amounts of data to different processes in the communicator. "
There was a problem hiding this comment.
I think I am missing something:
- i did the same change for alltoall
- this issue is not about collecting the datatypes, it is about fixing a problem with the current implementation. If we decide to collect all the datatypes (which i personally think should be done), it should be a separate issue; not this one.
There was a problem hiding this comment.
I agree it should be a separate issue. Just thought it worth a precautionary heads up before you work on the golang consumer of this data. Observation - the github interface is not so good a drawing attention to motivation for and decisions on scope of changes - or am I still a newbie looking in the wrong place?!
Fix #96 including a bug fix to correctly report the datatype size.