NHDaly
commented
1 year ago Note that of course one must be careful when interpreting the results of concurrently running profiles, since inference caused by invocation B will also show up in the profile for invocation A. But this is true of all of julia's profiling and timing techniques right now, so it's just something the developer must be aware of. (Compare it with @time from multiple tasks reporting allocations caused by other tasks, or reporting compilation time from other tasks.)
codecov[bot]
NOTE: It turns out this PR isn't mergeable as-is, since it depends on assumptions in Julia that aren't valid. Namely, we need https://github.com/JuliaLang/julia/pull/47258 and https://github.com/JuliaLang/julia/pull/47260 to avoid data races, but it turns out those don't make sense against julia master/1.9, because of recent changes to the inference lock.
At RAI, we have backported those two PRs against 1.8.2, and we're going to keep hacking on this branch to pull it into our internal build. We will follow-up with actual changes to julia 1.9 and SnoopCompile to try to make this work for real.
Following option 3 from this doc: https://docs.google.com/document/d/1JO43w3V1Cwtb2CxWf4sYfAB68xbkTrCMr973iaJOFwY/edit#
This is a WIP PR to support invoking
@snoopi_deepconcurrently and in parallel from multiple Tasks. (It currently requires https://github.com/JuliaLang/julia/pull/47258 to avoid data races, and https://github.com/JuliaLang/julia/pull/47260 to allow reentrant inference timing.)The idea is basically to share the global julia type inference profile buffer, rather than clear it at the start of every profile, keeping track of the start and stop index for each invocation of snoopi_deep. Then we also clear the beginning of the profile buffer whenever the oldest invocation ends, to ensure we aren't leaking memory.