-
Notifications
You must be signed in to change notification settings - Fork 12.7k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Revert lazy init #74716
Revert lazy init #74716
Conversation
r? @kennytm (rust_highfive has picked a reviewer for you, use r? to override) |
@eddyb -- we'll likely want to re-land the niche-filling PR, but I'd like to do so once we've resolved the current perf state fully just in case (though I now believe it to be perf-neutral). @bors try @rust-timer queue r? @nnethercote |
Awaiting bors try build completion |
⌛ Trying commit 6eaaff4 with merge 193b2f77c9463ed7378c20bad843a9031489e215... |
I expect the results here to roughly speaking be a perfect inverse of https://perf.rust-lang.org/compare.html?start=d3df8512d2c2afc6d2e7d8b5b951dd7f2ad77b02&end=cfade73820883adf654fe34fd6b0b03a99458a51, by the way. |
Hm, yeah, it is pretty unfortunate that an addition of std API which is not even used anywhere yet measurably regresses benchmarks. Proceduraly, do we want to block this until the compiler is fixed to deal with unused code faster? It might be reasonable in terms of end-user experience. But, this is really not fixable within the PR itself I think, which is not good either. In terms of overall story with compile-times, I am actually hoping that |
☀️ Try build successful - checks-actions, checks-azure |
Queued 193b2f77c9463ed7378c20bad843a9031489e215 with parent 9008693, future comparison URL. |
Oh, I think it's unclear whether we'll land this PR or not. I think probably yes but we may re-land the lazy init PR without fixing the compiler, just like we're somewhat likely to do with the backtrace stuff. I agree with you 100% that it is pretty annoying that just adding an API to std regresses performance for downstream crates despite lack of use. |
I am quite confused by these results. Let me summarize what we've run so far:
These results are pretty much nonsensical, so I attempted to reproduce them locally. I was unable to do so, at least on the helloworld and unify-linearly benchmarks -- I observe no performance difference in the original rollup. I then investigated what occurred on the collector between these two benchmarks. 2020-07-18 02:32:46.688688+00 is when d3df851 was collected. There are several differences that landed in that time range onto the collection server. We stopped building std on the collection server: rust-lang/rustc-perf@048360c. This was deployed on 2020-07-19 14:34:00, with 47ea6d9 being the first commit benchmarked since then. There were likely some other deploys and changes that I'll try to include later. I haven't yet figured out if we should assign all blame to the lack of std recompilation or what else, yet, though. It doesn't help that the collector benchmarks things out of order with respect to git history. This is a log of the commits and when they were benchmarked on the collection server. |
Finished benchmarking try commit (193b2f77c9463ed7378c20bad843a9031489e215): comparison url. Benchmarking this pull request likely means that it is perf-sensitive, so we're automatically marking it as not fit for rolling up. Please note that if the perf results are neutral, you should likely undo the rollup=never given below by specifying Importantly, though, if the results of this run are non-neutral do not roll this PR up -- it will mask other regressions or improvements in the roll up. @bors rollup=never |
Having thought about this some more, it seems quite likely that having std built locally was giving us some performance benefits, though I'm not sure why -- it seems like we tried pretty hard to get a 1-1 matching there. I have re-queued d3df851, the parent of the rollup in question to run. My hope is that once that finishes, we'll see that the perf results of the rollup (saved to https://perf.rust-lang.org/compare.html?start=d3df8512d2c2afc6d2e7d8b5b951dd7f2ad77b02-v1&end=7d31ffc1ac9e9ea356e896e63307168a64501b9d&stat=instructions:u) will be neutral, rather than a regression. If that's the case, then we'll probably just chalk this up to "everything's okay" and drop this revert. I'll also re-land #74069 since that'll mean it is indeed performance neutral (which is almost certain, given the results of its revert). If my theory is right, I expect that rebuilding d3df851 on perf will show that it "regressed" by approximately equal amounts as the rollup did. It would be good to then pull in @eddyb (since they wrote the original std-in-perf code) to try and see if we can come up with a reason for CI std to be a performance regression. (Maybe we can land something fixing CI std for everyone!). |
Please do a CI perf run first :) |
I am closing this PR as our investigation has led us to believe that the rollup in question, #74468, is not a regression and is indeed performance neutral. We've also updated the triage log to account for this and document what happened. |
…ercote Reland rust-lang#74069 Investigation in rust-lang#74716 has concluded that this PR is indeed not a regression (and in fact the rollup itself is not either). This reverts the revert in rust-lang#74611. r? @nnethercote cc @eddyb
It looks like the implementation of the perf-tests for rollups is subtly wrong: it used the tree of the rollup merge, which included all rollup merges prior to that (rather than just the benchmarked PR). #74611 resolved to a neutral perf result which means that it was not the regression in #74468. Furthermore, results from #74592 now give a fairly strong indication that the problem is indeed the lazy initialization PR (#72414).
I suspect that the root cause here is the same as the gimli PR -- adding a sizeable amount of code to libstd increases overheads across the board.
cc @nnethercote