[Perf] NAOT iOS HelloWorld size: 4 Regressions on 5/15/2024 8:03:38 AM

Run Information

Name	Value
Architecture	x64
OS	Mac OS X 10.18
Queue	IPhone
Baseline	3de068c2be232960f4cb3b233dd2b4014e2ce7c3
Compare	714a4420805ed53c311b05381c83c88894100fa9
Diff	Diff
Configs	CompilationMode:tiered, HybridGlobalization:true, iOSStripSymbols:true, RunKind:ios_scenarios, RuntimeType:nativeaot

Regressions in SOD - iOS HelloWorld NativeAOT .app Size nollvm nosymbols

Benchmark	Baseline	Test	Test/Base	Edge Detector
app Size nollvm nosymbols - Aggregate - No Extension ADX - Test Multi Config Graph	2.76 MB	2.89 MB	1.05	True
app Size nollvm nosymbols - SOD - iOS HelloWorld NativeAOT .app Size nollvm nosymbols ADX - Test Multi Config Graph	2.76 MB	2.89 MB	1.05	True
app Size nollvm nosymbols - pub ADX - Test Multi Config Graph	2.76 MB	2.89 MB	1.05	True
app Size nollvm nosymbols - pub/iosHelloWorld/HelloiOS.app/HelloiOS ADX - Test Multi Config Graph	2.76 MB	2.89 MB	1.05	True

Test Report

Repro

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

git clone https://github.com/dotnet/performance.git
py .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'SOD - iOS HelloWorld NativeAOT .app Size nollvm nosymbols*'

Details

SOD - iOS HelloWorld NativeAOT .app Size nollvm nosymbols

ETL Files

Histogram

JIT Disasms

Docs

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

@MichalStrehovsky

The range for regression is 778cc84...a71a85a. Could it be #102183 @MichalStrehovsky? ... but seems unlikely that adding assembly name would cause 5% regression in size...

@MichalStrehovsky

The range for regression is dotnet/runtime@778cc84...a71a85a. Could it be dotnet/runtime#102183 @MichalStrehovsky? ... but seems unlikely that adding assembly name would cause 5% regression in size...

If this is measuring size with symbols, it would be from that. Symbols should be stripped before measuring size though.

@MichalStrehovsky

The range for regression is dotnet/runtime@778cc84...a71a85a. Could it be dotnet/runtime#102183 @MichalStrehovsky? ... but seems unlikely that adding assembly name would cause 5% regression in size...

If this is measuring size with symbols, it would be from that. Symbols should be stripped before measuring size though.

This one should be without symbols (nosymbols in the issue header). With symbols we also saw regression of 6% (slightly bigger than this one) dotnet/perf-autofiling-issues#34369.

I've kicked off a size validation job in MichalStrehovsky/rt-sz#22 (results will be there in ~20 minutes) just to be sure, but that change should only affect symbols (and if it's from that, we should have seen an improvement in January when the regression my PR is fixing was introduced).

I've kicked off a size validation job in MichalStrehovsky/rt-sz#22 (results will be there in ~20 minutes) just to be sure, but that change should only affect symbols (and if it's from that, we should have seen an improvement in January when the regression my PR is fixing was introduced).

Looks like the size validation didn't show any regression. I wonder if it could be something mobile + NAOT specific as it doesn't show in you validation job yet the range doesn't contain anything else related to NAOT.

I've kicked off a size validation job in MichalStrehovsky/rt-sz#22 (results will be there in ~20 minutes) just to be sure, but that change should only affect symbols (and if it's from that, we should have seen an improvement in January when the regression my PR is fixing was introduced).

Looks like the size validation didn't show any regression. I wonder if it could be something mobile + NAOT specific as it doesn't show in you validation job yet the range doesn't contain anything else related to NAOT.

This should really only affect mangled names and mangled names are only used for symbols. Is it possible to get the binary that this is measuring? Just the after binary should be enough to rule out bad symbol stripping configuration.

@LoopedBard3

@LoopedBard3 do you think it is possible to get the binaries? I found the Perf build https://dev.azure.com/dnceng/internal/_build/results?buildId=2451389&view=logs&j=4f320d25-c1ea-5d1a-2cd7-ab8121a51286 but the Artifacts only contains binlogs.

@agocke

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

@ivanpovazan

fyi: @ivanpovazan

@LoopedBard3

@LoopedBard3 do you think it is possible to get the binaries? I found the Perf build https://dev.azure.com/dnceng/internal/_build/results?buildId=2451389&view=logs&j=4f320d25-c1ea-5d1a-2cd7-ab8121a51286 but the Artifacts only contains binlogs.

I think the binaries are right there:

Click the "Download iOS Sample App" step.

Downloading artifact iOSSampleAppNoSymbolsHybridGlobalizationTrue from: https://dev.azure.com/dnceng//_apis/resources/Containers/17496355?itemPath=iOSSampleAppNoSymbolsHybridGlobalizationTrue&isShallow=true&api-version=4.1-preview.4

Click that, follow the JSON to https://dev.azure.com/dnceng/_apis/resources/Containers/17496355?itemPath=iOSSampleAppNoSymbolsHybridGlobalizationTrue%2FiOSSampleAppNoSymbolsHybridGlobalizationTrue.zip.

And sure enough, the binary is not stripped properly. There should not be any S_P_CoreLib strings in a stripped binary, but there's lots of them. This is from #102183, but it's a benchmark setup issue; stripping doesn't do the right thing.

And sure enough, the binary is not stripped properly. There should not be any S_P_CoreLib strings in a stripped binary, but there's lots of them. This is from #102183, but it's a benchmark setup issue; stripping doesn't do the right thing.

Thanks Michal for pointing to the root cause!
We are applying the same stripping technique for both NativeAOT and MonoAOT to have "fair" comparison stripping only the local symbols (via: strip -x

runtime/src/tasks/AppleAppBuilder/Xcode.cs

Line 679 in 2d2d59c

    
           Utils.RunProcess(Logger, "strip", $"-no_code_signature_warning -x {appPath}/{filename}", workingDir: Path.GetDirectoryName(appPath));

) - this is a limitation with Mono (we have a tracking issue to improve this: #81530 so we can do full stripping).

That being said, in cases like this, we should check if the MAUI app also regressed (MAUI apps use full stripping):

if it did, then we most probably have an issue which we should investigate further
if it didn't, we should consider the regression with a grain of salt - limitation of how benchmark is setup

I took a look at our MAUI measurements and the regression for the range (84b3339...de6897b) containing this change is minimal (0.014 MB) + the MAUI version also changed dotnet/maui@562713c...13942bd so it is likely not even related.

@matouskozak

@matouskozak in that case I would suggest then closing this issue and link the problem we experienced here to #81530

Navigation Menu

Repository navigation

[Perf] NAOT iOS HelloWorld size: 4 Regressions on 5/15/2024 8:03:38 AM #104074

Run Information

Regressions in SOD - iOS HelloWorld NativeAOT .app Size nollvm nosymbols

Repro

SOD - iOS HelloWorld NativeAOT .app Size nollvm nosymbols

ETL Files

Histogram

JIT Disasms

Docs

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[Perf] NAOT iOS HelloWorld size: 4 Regressions on 5/15/2024 8:03:38 AM #104074

Description

Run Information

Regressions in SOD - iOS HelloWorld NativeAOT .app Size nollvm nosymbols

Repro

SOD - iOS HelloWorld NativeAOT .app Size nollvm nosymbols

ETL Files

Histogram

JIT Disasms

Docs

Activity

matouskozak commented on May 15, 2024

MichalStrehovsky commented on May 15, 2024

matouskozak commented on May 15, 2024

MichalStrehovsky commented on May 15, 2024

matouskozak commented on May 16, 2024

MichalStrehovsky commented on May 21, 2024

matouskozak commented on May 21, 2024

dotnet-policy-service commented on Jun 27, 2024

matouskozak commented on Jun 27, 2024

MichalStrehovsky commented on Jun 27, 2024

ivanpovazan commented on Jun 27, 2024

matouskozak commented on Jul 2, 2024

ivanpovazan commented on Jul 3, 2024

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions