[Mono] TE plaintext and json benchmarks regression

[fanyang-mono]

A recent change has caused regression for TE plaintext and json benchmarks running on Mono. The PR which caused this regression is #53450

cc: @imhameed

I couldn't figure out the best area label to add to this issue. If you have write-permissions please help me learn by adding exactly one area label.

[tannergooding]

@fanyang-mono, could you provide guidance on how to get the disassembly for Mono on these benchmarks, etc?

[tannergooding]

For what it's worth, this looks like a Mono bug and something its not handling correctly.
CoreCLR didn't regress and in fact looks like it has slightly better latency.

Given there were only two very minor mono file changes here: https://github.com/dotnet/runtime/pull/53450/files#diff-ce66e94ce2cdb8e54003a124a4057ce5d5dc5325ad32c9601a4662f8fa39367a

My guess is that Mono isn't handling some pattern or API which is getting handled by CoreCLR.

[fanyang-mono]

Could be. I will obtain the trace before and after the regression to see what changed for Mono.

[sebastienros]

@tannergooding the drop in the charts happens when we switched from net6.0 to net7.0 so it could have been from any commits for about 1-2 months of changes on main while we were still benchmarking net6.0. The improvements you are mentioning are for this 1-2 months period too, not specifically this change. @fanyang-mono identified the specific mono regression to two consecutive builds by running benchmarks manually on previous builds. It's possible that it also impacts the coreclr benchmarks, but other improvements in net7.0 still have a positive impact (ThreadPool improvements for instance).

@fanyang-mono can you share the runtime versions you used, and maybe do a run on coreclr for the same benchmarks between these two versions?

[tannergooding]

It's possible that it also impacts the coreclr benchmarks, but other improvements in net7.0 still have a positive impact (ThreadPool improvements for instance).

@sebastienros, that's certainly possible, but I also think its somewhat unlikely. We have the dotnet/performance repo and run daily per commit benchmarks with semi-weekly triage to identify regressions and improvements.

Before merging PRs like this we get diffs and there was nothing significant for RyuJIT in the frameworks, tests, or benchmarks libraries. There were likewise several improvements that were directly correlated to the PR and no flagged regressions from what we observed.

[naricc]

It looks like we are also seeing this regression in Microbenchmarks: dotnet/perf-autofiling-issues#2173 (that is a link to the automated issue). I don't think it makes sense for me to create another one, but using the microbenchmarks might be easier for validation of a fix.

[fanyang-mono]

@fanyang-mono can you share the runtime versions you used, and maybe do a run on coreclr for the same benchmarks between these two versions?

Runtime packages of this regression:
7.0.0-alpha.1.21480.11 -> good
7.0.0-alpha.1.21480.12 --> regression

[fanyang-mono]

It looks like we are also seeing this regression in Microbenchmarks: dotnet/perf-autofiling-issues#2173 (that is a link to the automated issue). I don't think it makes sense for me to create another one, but using the microbenchmarks might be easier for validation of a fix.

Among all the regressions filed for the above issue, these microbenchmarks showed regression exactly on Sept 30th.

https://pvscmdupload.blob.core.windows.net/autofilereport/autofilereports/11_08_2021/refs/heads/main_x64_ubuntu%2018.04_LLVM=false_MonoAOT=true_MonoInterpreter=false_Regression/System.Text.Json.Tests.Perf_Reader.html

https://pvscmdupload.blob.core.windows.net/autofilereport/autofilereports/11_08_2021/refs/heads/main_x64_ubuntu%2018.04_LLVM=false_MonoAOT=true_MonoInterpreter=false_Regression/System.Memory.Span(Char).html

[tannergooding]

Just noting a related PR went in: #60094

It would be interesting to see if Mono likewise sees a regression or any change here, particularly since these ones are showing 43% and 25% improvements in our perf triage from today: dotnet/perf-autofiling-issues#2339 and dotnet/perf-autofiling-issues#2346