[mono][size][Perf] iOS and Android size regression on 8/9/2024 8:00:01 AM #106265

performanceautofiler · 2024-08-09T08:00:02Z

Run Information

Name	Value
Architecture	x64
OS	Mac OS X 10.18
Queue	IPhone
Baseline	47ebcf3831a4a9b43d9abb19c21e4d66752e1d7c
Compare	68511fd27fe4055ce5203742998ba12019dfcbd4
Diff	Diff
Configs	CompilationMode:tiered, HybridGlobalization:true, iOSLlvmBuild:true, iOSStripSymbols:true, RunKind:ios_scenarios, RuntimeType:mono

Regressions in SOD - iOS HelloWorld Mono .app Size llvm nosymbols

Benchmark	Baseline	Test	Test/Base	Edge Detector
app Size llvm nosymbols - SOD - iOS HelloWorld Mono .app Size llvm nosymbols ADX - Test Multi Config Graph	13.38 MB	14.16 MB	1.06	True
app Size llvm nosymbols - pub/iosHelloWorld/HelloiOS.app/System.Console.dll ADX - Test Multi Config Graph	16.50 KB	17.00 KB	1.03	True
app Size llvm nosymbols - SOD - iOS HelloWorld Mono .app Size llvm nosymbols - Count ADX - Test Multi Config Graph	13.00 count	17.00 count	1.31	True
app Size llvm nosymbols - pub/iosHelloWorld/HelloiOS.app/System.Private.CoreLib.dll ADX - Test Multi Config Graph	1.19 MB	1.29 MB	1.09	True
app Size llvm nosymbols - Aggregate - .dll ADX - Test Multi Config Graph	1.21 MB	1.40 MB	1.16	True
app Size llvm nosymbols - Aggregate - .dll - Count ADX - Test Multi Config Graph	3.00 count	5.00 count	1.67	True
app Size llvm nosymbols - Aggregate - No Extension ADX - Test Multi Config Graph	10.29 MB	10.71 MB	1.04	True
app Size llvm nosymbols - pub - Count ADX - Test Multi Config Graph	13.00 count	17.00 count	1.31	True
app Size llvm nosymbols - pub/iosHelloWorld/HelloiOS.app/Program.deps.json ADX - Test Multi Config Graph	5.14 KB	5.44 KB	1.06	True
app Size llvm nosymbols - pub/iosHelloWorld/HelloiOS.app/HelloiOS ADX - Test Multi Config Graph	10.29 MB	10.71 MB	1.04	True
app Size llvm nosymbols - Aggregate - .aotdata ADX - Test Multi Config Graph	1.88 MB	2.04 MB	1.08	True
app Size llvm nosymbols - pub ADX - Test Multi Config Graph	13.38 MB	14.16 MB	1.06	True
app Size llvm nosymbols - Aggregate - .aotdata - Count ADX - Test Multi Config Graph	4.00 count	6.00 count	1.50	True
app Size llvm nosymbols - pub/iosHelloWorld/HelloiOS.app/aot-instances.aotdata ADX - Test Multi Config Graph	1.05 MB	1.13 MB	1.07	True
app Size llvm nosymbols - Aggregate - .json ADX - Test Multi Config Graph	6.62 KB	6.91 KB	1.04	True
app Size llvm nosymbols - pub/iosHelloWorld/HelloiOS.app/System.Private.CoreLib.aotdata ADX - Test Multi Config Graph	839.63 KB	847.38 KB	1.01	True

Test Report

Repro

General Docs link: https://github.com/dotnet/performance/blob/main/docs/benchmarking-workflow-dotnet-runtime.md

Repro Steps

Prerequisites (Build files either built locally or downloaded from payload above)

Libraries build extracted to runtime/artifacts or build instructions: Libraries README args: -subset libs+libs.tests -rc release -configuration Release -arch $RunArch -framework net8.0
CoreCLR product build extracted to runtime/artifacts/bin/coreclr/$RunOS.$RunArch.Release, build instructions: CoreCLR README args: -subset clr+libs -rc release -configuration Release -arch $RunArch -framework net8.0
Mono Runtime build extracted to runtime/artifacts/bin/mono/$RunOS.$RunArch.Release, build instructions: MONO README args: -arch $RunArch -os $RunOS -s mono+libs+host+packs -c Release
Dotnet SDK installed for dotnet commands
Running commands from the runtime folder

Linux

# Set $RunDir to the runtime directory
RunDir=`pwd`

# Set the OS, arch, and OSId
RunOS='linux'
RunOSId='linux'
RunArch='x64'

# Create mono dotnet
mkdir -p $RunDir/artifacts/dotnet-mono
$RunDir/build.sh -subset libs.pretest -configuration release -ci -arch $RunArch -testscope innerloop /p:RuntimeArtifactsPath=$RunDir/artifacts/bin/mono/$RunOS.$RunArch.Release /p:RuntimeFlavor=mono
cp $RunDir/artifacts/bin/runtime/net8.0-$RunOS-Release-$RunArch/* $RunDir/artifacts/bin/testhost/net8.0-$RunOS-Release-$RunArch/shared/Microsoft.NETCore.App/8.0.0 -rf
cp $RunDir/artifacts/bin/testhost/net8.0-$RunOS-Release-$RunArch/* $RunDir/artifacts/dotnet-mono -r
cp $RunDir/artifacts/bin/coreclr/$RunOS.$RunArch.Release/corerun $RunDir/artifacts/dotnet-mono/shared/Microsoft.NETCore.App/8.0.0/corerun

# Create Core Root
$RunDir/src/tests/build.sh release $RunArch generatelayoutonly /p:LibrariesConfiguration=Release

# Clone performance 
git clone --branch main --depth 1 --quiet https://github.com/dotnet/performance.git $RunDir/performance

# One line run:
python3 $RunDir/performance/scripts/benchmarks_ci.py --csproj $RunDir/performance/src/benchmarks/micro/MicroBenchmarks.csproj --incremental no --architecture $RunArch -f net8.0 --filter 'SOD - iOS HelloWorld Mono .app Size llvm nosymbols*' --bdn-artifacts $RunDir/artifacts/BenchmarkDotNet.Artifacts --bdn-arguments="--anyCategories Libraries Runtime  --category-exclusion-filter NoInterpreter NoMono --logBuildOutput --generateBinLog --corerun $RunDir/artifacts/dotnet-mono/shared/Microsoft.NETCore.App/8.0.0/corerun" 

# Individual Commands:
# Restore 
dotnet restore $RunDir/performance/src/benchmarks/micro/MicroBenchmarks.csproj --packages $RunDir/performance/artifacts/packages /p:UseSharedCompilation=false /p:BuildInParallel=false /m:1

# Build
dotnet build $RunDir/performance/src/benchmarks/micro/MicroBenchmarks.csproj --configuration Release --framework net8.0 --no-restore /p:NuGetPackageRoot=$RunDir/performance/artifacts/packages /p:UseSharedCompilation=false /p:BuildInParallel=false /m:1

# Run
dotnet run --project $RunDir/performance/src/benchmarks/micro/MicroBenchmarks.csproj --configuration Release --framework net8.0 --no-restore --no-build -- --filter 'SOD - iOS HelloWorld Mono .app Size llvm nosymbols*' --anyCategories Libraries Runtime " --category-exclusion-filter NoInterpreter NoMono --logBuildOutput --generateBinLog --corerun $RunDir/artifacts/dotnet-mono/shared/Microsoft.NETCore.App/8.0.0/corerun --artifacts $RunDir/artifacts/BenchmarkDotNet.Artifacts --packages $RunDir/performance/artifacts/packages

Windows

# Set $RunDir to the runtime directory
$RunDir="FullPathHere"

# Set the OS, arch, and OSId
RunOS='windows'
RunOSId='win'
RunArch='x64'

# Create mono dotnet
mkdir -p $RunDir/artifacts/dotnet-mono
$RunDir/build.sh -subset libs.pretest -configuration release -ci -arch $RunArch -testscope innerloop /p:RuntimeArtifactsPath=$RunDir\artifacts\bin\mono\$RunOS.$RunArch.Release /p:RuntimeFlavor=mono
xcopy $RunDir\artifacts\bin\runtime\net8.0-$RunOS-Release-$RunArch\ $RunDir\artifacts\bin\testhost\net8.0-$RunOS-Release-$RunArch\shared\Microsoft.NETCore.App\8.0.0\ /e /y
xcopy $RunDir\artifacts\bin\testhost\net8.0-$RunOS-Release-$RunArch\ $RunDir\artifacts\dotnet-mono\ /e /y
xcopy $RunDir\artifacts\bin\coreclr\$RunOS.$RunArch.Release\corerun $RunDir\artifacts\dotnet-mono\shared\Microsoft.NETCore.App\8.0.0\corerun /y

# Create Core Root
$RunDir\src\tests\build.cmd release $RunArch generatelayoutonly /p:LibrariesConfiguration=Release

# Clone performance 
git clone --branch main --depth 1 --quiet https://github.com/dotnet/performance.git $RunDir\performance

# One line run:
python3 $RunDir\performance\scripts\benchmarks_ci.py --csproj $RunDir\performance\src\benchmarks\micro\MicroBenchmarks.csproj --incremental no --architecture $RunArch -f net8.0 --filter 'SOD - iOS HelloWorld Mono .app Size llvm nosymbols*' --bdn-artifacts $RunDir\artifacts\BenchmarkDotNet.Artifacts --bdn-arguments="--anyCategories Libraries Runtime  --category-exclusion-filter NoInterpreter NoMono --logBuildOutput --generateBinLog --corerun $RunDir\artifacts\dotnet-mono\shared\Microsoft.NETCore.App\8.0.0\corerun.exe" 

# Individual Commands:
# Restore 
dotnet restore $RunDir\performance\src\benchmarks\micro\MicroBenchmarks.csproj --packages $RunDir\performance\artifacts\packages /p:UseSharedCompilation=false /p:BuildInParallel=false /m:1

# Build
dotnet build $RunDir\performance\src\benchmarks\micro\MicroBenchmarks.csproj --configuration Release --framework net8.0 --no-restore /p:NuGetPackageRoot=$RunDir\performance\artifacts\packages /p:UseSharedCompilation=false /p:BuildInParallel=false /m:1

# Run
dotnet run --project $RunDir\performance\src\benchmarks\micro\MicroBenchmarks.csproj --configuration Release --framework net8.0 --no-restore --no-build -- --filter 'SOD - iOS HelloWorld Mono .app Size llvm nosymbols*' --anyCategories Libraries Runtime " --category-exclusion-filter NoInterpreter NoMono --logBuildOutput --generateBinLog --corerun $RunDir\artifacts\dotnet-mono\shared\Microsoft.NETCore.App\8.0.0\corerun.exe --artifacts $RunDir\artifacts\BenchmarkDotNet.Artifacts --packages $RunDir\performance\artifacts\packages

SOD - iOS HelloWorld Mono .app Size llvm nosymbols

ETL Files

Histogram

JIT Disasms

Docs

Profiling workflow for dotnet/runtime repository
Benchmarking workflow for dotnet/runtime repository

Run Information

Name	Value
Architecture	x64
OS	Mac OS X 10.18
Queue	IPhone
Baseline	47ebcf3831a4a9b43d9abb19c21e4d66752e1d7c
Compare	68511fd27fe4055ce5203742998ba12019dfcbd4
Diff	Diff
Configs	CompilationMode:tiered, HybridGlobalization:true, iOSLlvmBuild:true, iOSStripSymbols:true, RunKind:ios_scenarios, RuntimeType:mono

Regressions in SOD - iOS HelloWorld Mono Zip Size llvm nosymbols

Benchmark	Baseline	Test	Test/Base	Edge Detector
SOD - iOS HelloWorld Mono Zip Size llvm nosymbols - pub ADX - Test Multi Config Graph	4.51 MB	4.77 MB	1.06	True
SOD - iOS HelloWorld Mono Zip Size llvm nosymbols - Aggregate - .zip ADX - Test Multi Config Graph	4.51 MB	4.77 MB	1.06	True
SOD - iOS HelloWorld Mono Zip Size llvm nosymbols - SOD - iOS HelloWorld Mono Zip Size llvm nosymbols ADX - Test Multi Config Graph	4.51 MB	4.77 MB	1.06	True
SOD - iOS HelloWorld Mono Zip Size llvm nosymbols - pub/iOSSampleApp.zip ADX - Test Multi Config Graph	4.51 MB	4.77 MB	1.06	True

Test Report

Repro

General Docs link: https://github.com/dotnet/performance/blob/main/docs/benchmarking-workflow-dotnet-runtime.md

Repro Steps

Prerequisites (Build files either built locally or downloaded from payload above)

Libraries build extracted to runtime/artifacts or build instructions: Libraries README args: -subset libs+libs.tests -rc release -configuration Release -arch $RunArch -framework net8.0
CoreCLR product build extracted to runtime/artifacts/bin/coreclr/$RunOS.$RunArch.Release, build instructions: CoreCLR README args: -subset clr+libs -rc release -configuration Release -arch $RunArch -framework net8.0
Mono Runtime build extracted to runtime/artifacts/bin/mono/$RunOS.$RunArch.Release, build instructions: MONO README args: -arch $RunArch -os $RunOS -s mono+libs+host+packs -c Release
Dotnet SDK installed for dotnet commands
Running commands from the runtime folder

Linux

# Set $RunDir to the runtime directory
RunDir=`pwd`

# Set the OS, arch, and OSId
RunOS='linux'
RunOSId='linux'
RunArch='x64'

# Create mono dotnet
mkdir -p $RunDir/artifacts/dotnet-mono
$RunDir/build.sh -subset libs.pretest -configuration release -ci -arch $RunArch -testscope innerloop /p:RuntimeArtifactsPath=$RunDir/artifacts/bin/mono/$RunOS.$RunArch.Release /p:RuntimeFlavor=mono
cp $RunDir/artifacts/bin/runtime/net8.0-$RunOS-Release-$RunArch/* $RunDir/artifacts/bin/testhost/net8.0-$RunOS-Release-$RunArch/shared/Microsoft.NETCore.App/8.0.0 -rf
cp $RunDir/artifacts/bin/testhost/net8.0-$RunOS-Release-$RunArch/* $RunDir/artifacts/dotnet-mono -r
cp $RunDir/artifacts/bin/coreclr/$RunOS.$RunArch.Release/corerun $RunDir/artifacts/dotnet-mono/shared/Microsoft.NETCore.App/8.0.0/corerun

# Create Core Root
$RunDir/src/tests/build.sh release $RunArch generatelayoutonly /p:LibrariesConfiguration=Release

# Clone performance 
git clone --branch main --depth 1 --quiet https://github.com/dotnet/performance.git $RunDir/performance

# One line run:
python3 $RunDir/performance/scripts/benchmarks_ci.py --csproj $RunDir/performance/src/benchmarks/micro/MicroBenchmarks.csproj --incremental no --architecture $RunArch -f net8.0 --filter 'SOD - iOS HelloWorld Mono Zip Size llvm nosymbols*' --bdn-artifacts $RunDir/artifacts/BenchmarkDotNet.Artifacts --bdn-arguments="--anyCategories Libraries Runtime  --category-exclusion-filter NoInterpreter NoMono --logBuildOutput --generateBinLog --corerun $RunDir/artifacts/dotnet-mono/shared/Microsoft.NETCore.App/8.0.0/corerun" 

# Individual Commands:
# Restore 
dotnet restore $RunDir/performance/src/benchmarks/micro/MicroBenchmarks.csproj --packages $RunDir/performance/artifacts/packages /p:UseSharedCompilation=false /p:BuildInParallel=false /m:1

# Build
dotnet build $RunDir/performance/src/benchmarks/micro/MicroBenchmarks.csproj --configuration Release --framework net8.0 --no-restore /p:NuGetPackageRoot=$RunDir/performance/artifacts/packages /p:UseSharedCompilation=false /p:BuildInParallel=false /m:1

# Run
dotnet run --project $RunDir/performance/src/benchmarks/micro/MicroBenchmarks.csproj --configuration Release --framework net8.0 --no-restore --no-build -- --filter 'SOD - iOS HelloWorld Mono Zip Size llvm nosymbols*' --anyCategories Libraries Runtime " --category-exclusion-filter NoInterpreter NoMono --logBuildOutput --generateBinLog --corerun $RunDir/artifacts/dotnet-mono/shared/Microsoft.NETCore.App/8.0.0/corerun --artifacts $RunDir/artifacts/BenchmarkDotNet.Artifacts --packages $RunDir/performance/artifacts/packages

Windows

# Set $RunDir to the runtime directory
$RunDir="FullPathHere"

# Set the OS, arch, and OSId
RunOS='windows'
RunOSId='win'
RunArch='x64'

# Create mono dotnet
mkdir -p $RunDir/artifacts/dotnet-mono
$RunDir/build.sh -subset libs.pretest -configuration release -ci -arch $RunArch -testscope innerloop /p:RuntimeArtifactsPath=$RunDir\artifacts\bin\mono\$RunOS.$RunArch.Release /p:RuntimeFlavor=mono
xcopy $RunDir\artifacts\bin\runtime\net8.0-$RunOS-Release-$RunArch\ $RunDir\artifacts\bin\testhost\net8.0-$RunOS-Release-$RunArch\shared\Microsoft.NETCore.App\8.0.0\ /e /y
xcopy $RunDir\artifacts\bin\testhost\net8.0-$RunOS-Release-$RunArch\ $RunDir\artifacts\dotnet-mono\ /e /y
xcopy $RunDir\artifacts\bin\coreclr\$RunOS.$RunArch.Release\corerun $RunDir\artifacts\dotnet-mono\shared\Microsoft.NETCore.App\8.0.0\corerun /y

# Create Core Root
$RunDir\src\tests\build.cmd release $RunArch generatelayoutonly /p:LibrariesConfiguration=Release

# Clone performance 
git clone --branch main --depth 1 --quiet https://github.com/dotnet/performance.git $RunDir\performance

# One line run:
python3 $RunDir\performance\scripts\benchmarks_ci.py --csproj $RunDir\performance\src\benchmarks\micro\MicroBenchmarks.csproj --incremental no --architecture $RunArch -f net8.0 --filter 'SOD - iOS HelloWorld Mono Zip Size llvm nosymbols*' --bdn-artifacts $RunDir\artifacts\BenchmarkDotNet.Artifacts --bdn-arguments="--anyCategories Libraries Runtime  --category-exclusion-filter NoInterpreter NoMono --logBuildOutput --generateBinLog --corerun $RunDir\artifacts\dotnet-mono\shared\Microsoft.NETCore.App\8.0.0\corerun.exe" 

# Individual Commands:
# Restore 
dotnet restore $RunDir\performance\src\benchmarks\micro\MicroBenchmarks.csproj --packages $RunDir\performance\artifacts\packages /p:UseSharedCompilation=false /p:BuildInParallel=false /m:1

# Build
dotnet build $RunDir\performance\src\benchmarks\micro\MicroBenchmarks.csproj --configuration Release --framework net8.0 --no-restore /p:NuGetPackageRoot=$RunDir\performance\artifacts\packages /p:UseSharedCompilation=false /p:BuildInParallel=false /m:1

# Run
dotnet run --project $RunDir\performance\src\benchmarks\micro\MicroBenchmarks.csproj --configuration Release --framework net8.0 --no-restore --no-build -- --filter 'SOD - iOS HelloWorld Mono Zip Size llvm nosymbols*' --anyCategories Libraries Runtime " --category-exclusion-filter NoInterpreter NoMono --logBuildOutput --generateBinLog --corerun $RunDir\artifacts\dotnet-mono\shared\Microsoft.NETCore.App\8.0.0\corerun.exe --artifacts $RunDir\artifacts\BenchmarkDotNet.Artifacts --packages $RunDir\performance\artifacts\packages

SOD - iOS HelloWorld Mono Zip Size llvm nosymbols

ETL Files

Histogram

JIT Disasms

Docs

Profiling workflow for dotnet/runtime repository
Benchmarking workflow for dotnet/runtime repository

The text was updated successfully, but these errors were encountered:

dotnet-policy-service · 2024-08-12T13:09:03Z

Tagging subscribers to this area: @directhex, @matouskozak
See info in area-owners.md if you want to be subscribed.

dotnet-policy-service · 2024-08-12T13:09:16Z

Tagging subscribers to 'os-ios': @vitek-karas, @kotlarmilos, @ivanpovazan, @steveisok, @akoeplinger
See info in area-owners.md if you want to be subscribed.

dotnet-policy-service · 2024-08-12T13:09:22Z

Tagging subscribers to 'size-reduction': @eerhardt, @SamMonoRT, @marek-safar
See info in area-owners.md if you want to be subscribed.

matouskozak · 2024-08-12T13:10:17Z

The range of regression is 105bf4d...68511fd. ~~@kg do you think that #106080 could have affected this? The range is quite short and this looks to be the only Mono-related commit.~~

I did manual verification and the causing commit is #106014 @noahfalk

The same range is causing ~100kB size regression on Android sample app as well

dotnet-policy-service · 2024-08-12T13:14:21Z

Tagging subscribers to 'arch-android': @vitek-karas, @simonrozsival, @steveisok, @akoeplinger
See info in area-owners.md if you want to be subscribed.

matouskozak · 2024-08-15T07:45:54Z

@noahfalk @tarekgh the #106415 didn't have impact on Android/iOS size as expected

fyi: @vitek-karas

noahfalk · 2024-08-15T07:53:54Z

What diagnostics are available to determine the APIs that are no longer being trimmed and the rooting path preventing their removal? I am unfamiliar with how these size issues should be diagnosed.

vitek-karas · 2024-08-15T08:30:35Z

If you know which API was left in the binary and shouldn't have been then there are several ways to tell why it was kept:

ILSpy or similar will be able to find the callsites and such, lot of the times this is enough to figure out what's going on.
Trimmer can output a special log file which is basically the graph of dependencies it followed and we have tools to look into that (https://github.com/dotnet/runtime/tree/main/src/coreclr/tools/aot/DependencyGraphViewer)

To figure out which API is there and shouldn't be is typically a much more difficult problem and requires domain knowledge. For NativeAOT we can use https://github.com/MichalStrehovsky/sizoscope to diff two binaries, but trimmer doesn't support that yet.

noahfalk · 2024-08-15T08:43:02Z

I can make guesses that the APIs causing the size increase are either EventSource APIs or Meter APIs. I don't know how to verify those guesses. I'm not sure what app this size measuring automation is building and how I would reproduce it locally.

vitek-karas · 2024-08-15T08:45:06Z

I'm trying to get the IL of the trimmed app - before/after - then we'll have something to work with.

vitek-karas · 2024-08-15T10:52:27Z

So the comparison is pretty clear:
Before the app only has S.P.CoreLib and S.Console. After the app also has S.Diagnostics.DiagnosticSource and S.Collections.Concurrent. which add ~100KB of IL.

I'll look into it some more this happens and where it came from.

vitek-karas · 2024-08-15T19:10:11Z

So I found first bug - the substitution doesn't define default value, so if the feature switch is not defined at all (which is the default), no substitution happens. We'll have to add featuredefaultvalue to the XML to fix this. Unfortunately even if I work around this I still can't get the app to smaller size. I can see the substitution happening correctly, but it's as if something else pulls in the code as well - still looking (first time doing some real work on a mac ... and it hurts :-))

noahfalk · 2024-08-15T21:09:25Z

but it's as if something else pulls in the code as well

As far as I know the only link from S.P.CoreLib -> DiagnosticSource is the Type.GetType() call here. That call is supposed to be guarded by both the IsMeterSupported and EventSource.IsSupported checks, but of course maybe it isn't working as intended. If you want to you could comment out the call to PreregisterEventProviders at 3843 + comment out all of GetMetricsEventSource() and rebuild to confirm that is the code that is causes the dependency.

vitek-karas · 2024-08-16T12:17:14Z

OK - I think I understand this fully. The behavior is the same on Windows as well BTW.

If I compare P6 and RC1 - win-x64 NativeAOT (so not just normal trimming)

P6 - 1274368 bytes
RC1 - 1349120 bytes

So we regressed the size of console hello world by 75KB, so by almost 6%.

The reason is the same, we now include parts of System.Diagnostics.DiagnosticSource.dll and System.Collections.Concurrent.dll, before the change these assemblies were completely trimmed away.

Defaults

Currently both System.Diagnostics.Tracing.EventSource.IsSupported and System.Diagnostics.Metrics.Meter.IsSupported default to on for Windows (and for iOS as well).
As such this regression is currently "by design".

Trimming problem

If I turn off System.Diagnostics.Metrics.Meter.IsSupported (but leave EventSource on), the trimming doesn't really work. With the recent feature switch fix, we correctly trim away the callsite, but the code behind it is still kept. This is because EventSource has DAM.All annotation on it and if the EventSource.IsSupported is enabled, there's a code which makes use of it EnsureDescriptorsInitialized - so the trimmer must keep everything on EventSource - this brings in the GetMetricsEventSource method which in turn brings in the additional assemblies. So even though at runtime this code is unreachable, trimmer is instructed to keep it anyway.

To fix this, the cleanest solution would be to create a separate type from EventSource, and move the code which pulls in the metrics stuff there. That will avoid the DAM.All pulling in that code even if there's no callsite.

Side note: The fact that EventSource has DAM.All on it means, that if EventSource.IsSupported is on (which is the default basically everywhere), everything on that class is preserved - not sure if there are other trim opportunities we've missed because of this.

Acceptable size regression?

The second interesting question is - Do we need/want to enable metrics by default - on Windows console, on iOS, on other targets. I don't know what functionality it provides and if it's something every app should have enabled by default. @noahfalk maybe you could point us to some reading on this, so that we can decide if it makes sense to enable it for mobile targets. I'll leave the Windows/Linux/macOS targets up to you :-)

Measuring the right stuff

AFAIK we're not aware of NativeAOT size regression due to this change - right @MichalStrehovsky . That raises the question which infra should have caught it and why it didn't.

We'll have to live with this for RC1, but we should try to answer all these questions for RC2.

MichalStrehovsky · 2024-08-16T13:15:02Z

Good catch with the EnsureDescriptorsInitialized being considered a target of reflection (and therefore not removed at all) because EventSource reflects on its methods and methods of the descendants.

I think moving that method out into a separate type sounds like a good fix.

I'm not seeing the size regression for native AOT hello world on Windows. I see 1,104,384 bytes with P6 on a dotnet new console --aot and 1,099,264 bytes with the latest main.

We default EventSourceSupport to false on native AOT. Some SDKs like ASP.NET override the default to true.

I do see regression on a console hello world with EventSourceSupport overriden to true (from 2,842,624 bytes to 3,149,824 bytes). There seems to be tons of new metrics stuff. The regression was smaller for ASP.NET that was already calling into a lot of the metrics (the regression was about 30 kB because the only new "expensive" API getting called was Type.GetType. Our instrumentation is not very lightweight.

It would be nice if someone who understands metrics to confirm the 300 kB regression is expected. Download and unpack this zip:
beforeafter.zip

Then a Windows machine, run dotnet tool install sizoscope --global. Then run sizoscope and drag and drop before/testhello.mstat into the UI. Then while holding Alt key, drag and drop after/testhello.mstat. You'll see a UI that shows new things in a tree on the right side. You can double click things to see why the new thing was included. For example:

Notice the nodes in a different color only exists in "compare", not in "baseline". So for example in this case a new Dictionary<__Canon,bool> can be attributed to InitializeDefaultEventSources.

noahfalk · 2024-08-16T22:26:36Z

Thanks for investigating!

To fix this, the cleanest solution would be to create a separate type from EventSource, and move the code which pulls in the metrics stuff there. That will avoid the DAM.All pulling in that code even if there's no callsite.

Can the separate type be an inner class or will DAM.All pull that in too? I'm fine to do it as a top-level class if necessary, but inner class seemed a little nicer if the only point of the class is to get the trimmer behavior we want.

@noahfalk maybe you could point us to some reading on this, so that we can decide if it makes sense to enable it for mobile targets. I'll leave the Windows/Linux/macOS targets up to you :-)

These docs should give a good idea what scenarios you lose out on if you disable Meter:

How users observe metrics - https://learn.microsoft.com/en-us/dotnet/core/diagnostics/metrics-collection
How users create their own custom metrics - https://learn.microsoft.com/en-us/dotnet/core/diagnostics/metrics-instrumentation
Built-in metrics - https://learn.microsoft.com/en-us/dotnet/core/diagnostics/built-in-metrics

Metrics are most prevalent in web-services, but they can also be useful for ad-hoc performance investigations in all types of apps. In terms of defaults, my initial thought is that turning on/off EventSource seems like a better lever to use than Meter. If you are planning to turn Meter off by default and leave EventSource on I'd be curious to better understand the rationale.

It would be nice if someone who understands metrics to confirm the 300 kB regression is expected.

According to sizoscope it showed me the regression as 250.4 rather than 300, but aside from that yeah, the new dependencies from metrics seemed right.

noahfalk · 2024-08-16T22:52:13Z

Our instrumentation is not very lightweight.

Yeah, size optimization for instrumentation has never been a goal we've prioritized. The feedback that I see has been mostly about adding new capabilities, integration, standardization and performance at runtime. In many cases I suspect the perf at runtime has direct tradeoffs against code size as the fastest/lowest alloc implementations tend to be heavier on custom structs and generic instantiations of code to process them. I'm happy to chat more about it any time.

MichalStrehovsky · 2024-08-19T08:13:34Z

According to sizoscope it showed me the regression as 250.4 rather than 300, but aside from that yeah, the new dependencies from metrics seemed right.

Sizoscope doesn't have exact accounting of things - it pretty much always underreports because some bytes of the output are too difficult to attribute but those bytes are generally related to things that are visible in the UI.

Thank you for checking!

MichalStrehovsky · 2024-08-19T08:18:07Z

Can the separate type be an inner class or will DAM.All pull that in too?

This would need to be a separate class - .All includes nested types for better or worse (in this case for better, because event source reflects at nested types too).

runtime/src/libraries/System.Private.CoreLib/src/System/Diagnostics/Tracing/EventSource.cs

Lines 3141 to 3159 in 0fca85f

    
           // Collect task, opcode, keyword and channel information 
        
           foreach (string providerEnumKind in (ReadOnlySpan<string>)["Keywords", "Tasks", "Opcodes"]) 
        
           { 
        
               Type? nestedType = eventSourceType.GetNestedType(providerEnumKind); 
        
               if (nestedType != null) 
        
               { 
        
                   if (eventSourceType.IsAbstract) 
        
                   { 
        
                       manifest.ManifestError(SR.Format(SR.EventSource_AbstractMustNotDeclareKTOC, nestedType.Name)); 
        
                   } 
        
                   else 
        
                   { 
        
                       foreach (FieldInfo staticField in nestedType.GetFields(BindingFlags.DeclaredOnly | BindingFlags.Public | BindingFlags.NonPublic | BindingFlags.Static)) 
        
                       { 
        
                           AddProviderEnumKind(manifest, staticField, providerEnumKind); 
        
                       } 
        
                   } 
        
               } 
        
           }

noahfalk · 2024-08-19T08:21:53Z

Haha, I didn't even realize inner class reflection was in EventSource's bag of tricks. I'll have to tuck that tid-bit away somewhere. Thanks for the info! I'll put together another change that moves the code into a separate top-level internal class soon, hopefully tomorrow.

Recent changes to EventSource startup caused the IL trimmer to include portions of the DiagnosticSource assembly. Adding the IsMeterSupported feature check wasn't sufficient because EventSource also roots all of its methods via a DynamicallyAccessedMembers attribute. To ensure that the new methods can be removed when needed I refactored them into a separate EventSourceInitHelper class that won't be rooted by the existing DAM attribute. When EventSouce.IsSupported is false I'd expect the entire EventSourceInitHelper class to be unreachable. If only EventSource.IsMeterSupported is false then I'd expect PreregisterEventProviders and GetMetricsEventSource() to be unreachable but the rest of the class will remain. Hopefully this really fixes dotnet#106265 this time.

Recent changes to EventSource startup caused the IL trimmer to include portions of the DiagnosticSource assembly. Adding the IsMeterSupported feature check wasn't sufficient because EventSource also roots all of its methods via a DynamicallyAccessedMembers attribute. To ensure that the new methods can be removed when needed I refactored them into a separate EventSourceInitHelper class that won't be rooted by the existing DAM attribute. When EventSouce.IsSupported is false I'd expect the entire EventSourceInitHelper class to be unreachable. If only EventSource.IsMeterSupported is false then I'd expect PreregisterEventProviders and GetMetricsEventSource() to be unreachable but the rest of the class will remain. Hopefully this really fixes #106265 this time.

* Fix trimming for DiagnosticSource Recent changes to EventSource startup caused the IL trimmer to include portions of the DiagnosticSource assembly. Adding the IsMeterSupported feature check wasn't sufficient because EventSource also roots all of its methods via a DynamicallyAccessedMembers attribute. To ensure that the new methods can be removed when needed I refactored them into a separate EventSourceInitHelper class that won't be rooted by the existing DAM attribute. When EventSouce.IsSupported is false I'd expect the entire EventSourceInitHelper class to be unreachable. If only EventSource.IsMeterSupported is false then I'd expect PreregisterEventProviders and GetMetricsEventSource() to be unreachable but the rest of the class will remain. Hopefully this really fixes #106265 this time. * PR feedback --------- Co-authored-by: Noah Falk <noahfalk@microsoft.com> Co-authored-by: Jeff Schwartz <jeffschw@microsoft.com>

performanceautofiler bot added arch-x64 runtime-mono specific to the Mono runtime untriaged New issue has not been triaged by the area owner labels Aug 9, 2024

performanceautofiler bot mentioned this issue Aug 9, 2024

[SENTINEL] Autofile run complete at 8/9/2024 8:00:05 AM. 1 issues filed. dotnet/perf-autofiling-issues#39709

Closed

matouskozak mentioned this issue Aug 12, 2024

[Perf] Windows/x64: 19 Regressions on 8/11/2024 7:57:33 AM dotnet/perf-autofiling-issues#39731

Closed

matouskozak removed the untriaged New issue has not been triaged by the area owner label Aug 12, 2024

matouskozak changed the title ~~[Perf] Windows/x64: 20 Regressions on 8/9/2024 8:00:01 AM~~ [iOS][size][Perf] iOS size regression on 8/9/2024 8:00:01 AM Aug 12, 2024

dotnet-issue-labeler bot added the area-Infrastructure-mono label Aug 12, 2024

matouskozak transferred this issue from dotnet/perf-autofiling-issues Aug 12, 2024

dotnet-policy-service bot added the untriaged New issue has not been triaged by the area owner label Aug 12, 2024

matouskozak added arch-arm64 os-ios Apple iOS size-reduction Issues impacting final app size primary for size sensitive workloads and removed arch-x64 labels Aug 12, 2024

matouskozak changed the title ~~[iOS][size][Perf] iOS size regression on 8/9/2024 8:00:01 AM~~ [mono][size][Perf] iOS and Android size regression on 8/9/2024 8:00:01 AM Aug 12, 2024

matouskozak added the os-android label Aug 12, 2024

matouskozak mentioned this issue Aug 12, 2024

.NET 9 per-preview performance overview #98209

Open

tarekgh added this to the 9.0.0 milestone Aug 15, 2024

noahfalk mentioned this issue Aug 20, 2024

Fix trimming for DiagnosticSource #106680

Merged

dotnet-policy-service bot added the in-pr There is an active PR which will close this issue when it is merged label Aug 20, 2024

jeffschwMSFT removed the untriaged New issue has not been triaged by the area owner label Aug 20, 2024

noahfalk closed this as completed in #106680 Aug 22, 2024

noahfalk closed this as completed in 6177a9f Aug 22, 2024

noahfalk mentioned this issue Aug 22, 2024

[release/9.0] Fix trimming for DiagnosticSource #106842

Merged

4 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[mono][size][Perf] iOS and Android size regression on 8/9/2024 8:00:01 AM #106265

[mono][size][Perf] iOS and Android size regression on 8/9/2024 8:00:01 AM #106265

performanceautofiler bot commented Aug 9, 2024

Prerequisites (Build files either built locally or downloaded from payload above)

SOD - iOS HelloWorld Mono .app Size llvm nosymbols

ETL Files

Histogram

JIT Disasms

Docs

Prerequisites (Build files either built locally or downloaded from payload above)

SOD - iOS HelloWorld Mono Zip Size llvm nosymbols

ETL Files

Histogram

JIT Disasms

Docs

dotnet-policy-service bot commented Aug 12, 2024

dotnet-policy-service bot commented Aug 12, 2024

dotnet-policy-service bot commented Aug 12, 2024

matouskozak commented Aug 12, 2024 •

edited

Loading

dotnet-policy-service bot commented Aug 12, 2024

matouskozak commented Aug 15, 2024

noahfalk commented Aug 15, 2024

vitek-karas commented Aug 15, 2024

noahfalk commented Aug 15, 2024 •

edited

Loading

vitek-karas commented Aug 15, 2024

vitek-karas commented Aug 15, 2024

vitek-karas commented Aug 15, 2024

noahfalk commented Aug 15, 2024

vitek-karas commented Aug 16, 2024

MichalStrehovsky commented Aug 16, 2024

noahfalk commented Aug 16, 2024

noahfalk commented Aug 16, 2024

MichalStrehovsky commented Aug 19, 2024

MichalStrehovsky commented Aug 19, 2024

noahfalk commented Aug 19, 2024

[mono][size][Perf] iOS and Android size regression on 8/9/2024 8:00:01 AM #106265

[mono][size][Perf] iOS and Android size regression on 8/9/2024 8:00:01 AM #106265

Comments

performanceautofiler bot commented Aug 9, 2024

Run Information

Regressions in SOD - iOS HelloWorld Mono .app Size llvm nosymbols

Repro

Prerequisites (Build files either built locally or downloaded from payload above)

SOD - iOS HelloWorld Mono .app Size llvm nosymbols

ETL Files

Histogram

JIT Disasms

Docs

Run Information

Regressions in SOD - iOS HelloWorld Mono Zip Size llvm nosymbols

Repro

Prerequisites (Build files either built locally or downloaded from payload above)

SOD - iOS HelloWorld Mono Zip Size llvm nosymbols

ETL Files

Histogram

JIT Disasms

Docs

dotnet-policy-service bot commented Aug 12, 2024

dotnet-policy-service bot commented Aug 12, 2024

dotnet-policy-service bot commented Aug 12, 2024

matouskozak commented Aug 12, 2024 • edited Loading

dotnet-policy-service bot commented Aug 12, 2024

matouskozak commented Aug 15, 2024

noahfalk commented Aug 15, 2024

vitek-karas commented Aug 15, 2024

noahfalk commented Aug 15, 2024 • edited Loading

vitek-karas commented Aug 15, 2024

vitek-karas commented Aug 15, 2024

vitek-karas commented Aug 15, 2024

noahfalk commented Aug 15, 2024

vitek-karas commented Aug 16, 2024

Defaults

Trimming problem

Acceptable size regression?

Measuring the right stuff

MichalStrehovsky commented Aug 16, 2024

noahfalk commented Aug 16, 2024

noahfalk commented Aug 16, 2024

MichalStrehovsky commented Aug 19, 2024

MichalStrehovsky commented Aug 19, 2024

noahfalk commented Aug 19, 2024

matouskozak commented Aug 12, 2024 •

edited

Loading

noahfalk commented Aug 15, 2024 •

edited

Loading