Summary:
This is actually something I discovered a while ago with the wall of serotonin. It was really easy for large scale runs to get bottlenecked on disk access. I have a hack in the working files of that machine to use `/dev/shm`, but I figured I should formalize and actually make a respectable utility.
I also added a param to tweak the run cadence and print when a CorePool is created; these are just to make the CI logs a bit nicer. (A printout each second on a 40 minute CI job is a bit much...)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/56711
Reviewed By: agolynski
Differential Revision: D28392248
Pulled By: robieta
fbshipit-source-id: b6aa7445c488d8e4ab9d4b31ab18df4e12783d8f
Summary:
This PR add a `--mode` flag and a script to collect microbenchmarks in a single JSON file. I also added a version check since benchmarks are expected to evolve; this also turned up a determinism bug in `init_from_variants`. (`set` is not ordered, unlike `dict`)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/55428
Test Plan:
Run in CI
CC: ngimel wconstab ezyang bhosmer
Reviewed By: mruberry
Differential Revision: D27775284
Pulled By: robieta
fbshipit-source-id: c8c338fedbfb2860df207fe204212a0121ecb006
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/54652
This PR adds a fairly robust runner for the instruction count microbenchmarks. Key features are:
* Timeout and retry. (In rare cases, Callgrind will hang under heavy load.)
* Robust error handling and keyboard interrupt support.
* Benchmarks are pinned to cores. (Wall times still won't be great, but it's something.)
* Progress printouts, including a rough ETA.
Test Plan: Imported from OSS
Reviewed By: pbelevich
Differential Revision: D27537823
Pulled By: robieta
fbshipit-source-id: 699ac907281d28bf7ffa08594253716ca40204ba
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/54651
This PR fleshes out the benchmarks to everything I could come up with. (166 individual cases when all is said and done.) If there's anything you feel warrants a spot in CI that I've missed, by all means let me know.
Test Plan: Imported from OSS
Reviewed By: pbelevich
Differential Revision: D27537824
Pulled By: robieta
fbshipit-source-id: 3819e8fec2131c6b5f29f5099cd41e79131bed90
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53296
Part 1 of the instruction count microbenchmarks. This PR is focused on benchmark definition machinery. (Though you can run `main.py` to see it in action.) A summary of the system is given in the README.
Test Plan: Imported from OSS
Reviewed By: ngimel
Differential Revision: D26907092
Pulled By: robieta
fbshipit-source-id: 0f61457b3ce89aa59a06bf1f0e7a74ccdbf17090
