FIT

Fault Injection Tool:

Using Python to parse a c file and inject faults into it according to a specified probability distribution

Features

1. Injects soft errors into a program (currently into the mlp.cc file provided)

• looks for innermost for-loop
• within this for-loop, it perturbs every statement that uses the += operation

2. Produces an output file which it can compile and run
3. Extracts accuracy data and writes it to a results file

Requirements

1. Linux (Tested on XUbuntu 14.04)
2. g++
3. python (Tested with python3)

Getting Started

Note: There are some outdated files in this directory for the purpose of learning. These files are fit_merge.py, i.cc, test.cc, and test.cfg. The fit_merge.py file will actually create an output file that contains the error code directly inside. This was changed in the newer version.

Without customization, this code can be run to inject errors into mlp.cc by running the commandpython3 fit.py mlp.cfg in terminal.

To customize this code, edit the configuration file, mlp.cfg. The format of the config file follows Python's configparser library. Functions should be separated by spaces, and for every function there must be a corresponding percent. This percent refers to what percentage of the total execution time of the program is spent in the corresponding function.

test.cc and test.cfg do not work with fit.py for a specific reason. In order to obtain the accuracy results, the current design of the tool expects the original file to contain a statement resembling printf("%d : %d\n", counter, test_size); Only through such a statement can the tool determine the accuracy percentage. This is something to keep in mind when trying to make the tool work for a program other than mlp.cc.

Future Considerations

1. Accept a list of potential non-crucial functions to perturb
   • create configuration files for each possible combination of functions
   • analyze which configuration file is ideal
2. Accept multiple error rates and gather the accuracy results for each rate
3. Introduce bounds checking
4. Automate time evaluation process of determining how long is spent executing each portion of non-crucial code
5. Account for type of value to be perturbed and introduce error accordingly (currently assuming type double)
6. Account for a variety of function types to be perturbed (currently only looking at innermost for-loops and += operations)