Stream: cranelift

Topic: FYI: sightglass no longer has per-process overheads

view this post on Zulip fitzgen (he/him) (Aug 11 2022 at 20:39):

FYI for folks who use sightglass for benchmarking cranelift changes: as of https://github.com/bytecodealliance/sightglass/pull/193, it no longer has significant per-process overheads. So if you were previously doing something like --processes 1 --iterations 100, you can now safely do something like --processes 10 --iterations-per-process 10 (which happens to be the default) without benchmarking taking a bunch longer than it otherwise would.

cc @Chris Fallin @Jamey Sharp @Trevor F. Smith @Alex Crichton

Disable wall time measurement in "wall-cycles" measure; only measure cycles by fitzgen · Pull Request #193 · bytecodealliance/sightglass
Rather than the precision crate, which is way too over-eager with its calibration, and often takes 5 seconds to calibrate. This in turn was making multi-process benchmark measurements way too slow ...

view this post on Zulip Chris Fallin (Aug 11 2022 at 20:42):

awesome, thanks for doing this work!

Last updated: Nov 22 2024 at 17:03 UTC