search

IO500 / io500

IO500 Storage Benchmark source code
MIT License
95 stars 30 forks source link

Support for pausing individual benchmarks. #55

Closed JulianKunkel closed 2 years ago

JulianKunkel commented 2 years ago

This may be used for testing of e.g., tiered storage.

Setting the pause-dir directory option, the rank0 will check before starting each phase if a file exists in the pause dir. If it does, it will not start. For debugging, it will print a message for each phase into the summary.txt file: ; Checking for pause file ./pause/mdtest-hard-write

gflofst commented 2 years ago

This seems like a clunky way to do this. From a scoring standpoint, I am not seeing something that disqualifies something that pauses and no recognition in the output that the pause occurs and for how long. If we are going to add this feature, we should do it in a way that recognizes this is part of the official IO500 suite and the potential on scoring.

gflofst commented 2 years ago

I believe we agreed that it would invalidate. My request is that this pull request lean in on invalidation my being explicit in the output and ideally using something that does not accidentally make a run invalid (forgot to delete the directory, but the scripts all looked good). A message at the start would also be a good indicator saying WARNING: THIS RUN WILL NOT BE VALID FOR COMPETITION AS A PAUSE WILL OCCUR

JulianKunkel commented 2 years ago

The patch already included the information regarding a pause into the summary.txt file, hence it is transparent what is happening. I included that the run is invalidated. There might be cases where the run is still valid, though...