Mono perf regression based on TechEmpower PlaintextPlatform benchmark

[fanyang-mono]

Recently, there is a ~5% regression on TechEmpower PlaintextPlatform benchmark for all 3 different flavors of mono (JIT, LLVM JIT, LLVM AOT)

Based on bisect, the regression was introduced by 778be98

For mono JIT, with the commit before this one, RPS (requests per second) was 9,173,356; with this commit, RPS was 8,709,109. For LLVM JIT and LLVM AOT, the regression percentage was similar. For LLVM AOT, the number went from 10,244,296 to 9,795,509.

To get the results from your own machine:

1. Clone https://github.com/aspnet/Benchmarks
2. cd Benchmarks/src/BenchmarksDriver2
3. dotnet run -- --config ../BenchmarksApps/Kestrel/PlatformBenchmarks/platform.benchmarks.yml --scenario plaintext --profile aspnet-citrine-lin --application.selfContained true --application.runtimeVersion 5.0.0-preview.8.20324.2 -i 3 --application.options.outputFiles path_to/System.Private.CoreLib.dll --application.options.outputFiles path_to/libcoreclr.so

• Note that System.Private.CoreLib.dll and libcoreclr.so should comes from linux build. They usually live at
  artifacts/obj/mono/Linux.x64.Release/mono/mini/.libs/libmonosgen-2.0.so (needs to be rename to libcoreclr.so)
  artifacts/bin/mono/Linux.x64.Release/System.Private.CoreLib.dll

/cc @tannergooding @imhameed

[ghost]

Tagging subscribers to this area: @lewurm
Notify danmosemsft if you want to be subscribed.

[tannergooding]

Mono fully supports Vector<T>, correct? If so, this sounds like a dead-code optimization issue

[danmoseley]

How can this commit affect Mono benchmark performance? It's a change to CoreCLR JIT, and associated tests.

[tannergooding]

It was a change to the library code, to ensure that it properly throws for unsupported T. So it would also impact Mono

[danmoseley]

Oh! My bad. I did not scroll enough.

[imhameed]

I'll look into this later today.

[fanyang-mono]

With Imran's recent PR #39322, RPS for mono JIT went from 8,351,923 to 8,409,044.

[SamMonoRT]

We haven't had regular runs in the perf labs since Jun 25th as they are upgrading the driver system and can't be certain if there were other regressions creeping in while the lab was offline(Mono runs not being executed). We haven't quite achieved the same perf numbers as before, and are waiting on the new runs to be enabled in perf labs. if those numbers from newer runs are trending in positive direction we can close this issue. @fanyang-mono @imhameed

[vargaz]

Would it be possible to rerun the tests after this:
fa72c75

[SamMonoRT]

@vargaz - Yes, luckily the official runs(atleast for JIT) have just restarted in the labs again. We should have newer set of results at end of the day.

[SamMonoRT]

@vargaz new numbers from last night's run are published. For plaintextplatform benchmark, MonoVM- Jit config had ~3% increase. I have not pinpointed the exact change which attributed to the increase, but your change in fa72c75 could surely have had an effect.
8.828M reqs/sec June 26th
9.081M reqs/sec June 28th (which should have your change merged)

[SamMonoRT]

We can close this issue.