[Perf] Linux/arm64: 4 Regressions on 5/3/2024 3:54:35 PM

[performanceautofiler]

Run Information

Name	Value
Architecture	arm64
OS	ubuntu 22.04
Queue	AmpereUbuntu
Baseline	fc76b1cac3f02cc9729f6682d6850fd7982e9fe5
Compare	358b0a4d350a6c72ccc825a0ac668620f849365a
Diff	Diff
Configs	CompilationMode:tiered, RunKind:micro

Regressions in System.Text.Json.Tests.Utf8JsonReaderCommentsTests

Benchmark	Baseline	Test	Test/Base	Test Quality	Edge Detector	Baseline IR	Compare IR	IR Ratio
Utf8JsonReaderCommentParsing - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	78.93 ns	100.11 ns	1.27	0.13	False
Utf8JsonReaderCommentParsing - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	80.90 ns	93.24 ns	1.15	0.34	False
Utf8JsonReaderCommentParsing - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	94.71 ns	125.39 ns	1.32	0.48	False
Utf8JsonReaderCommentParsing - Duration of single invocation 📝 - Benchmark Source ADX - Test Multi Config Graph	81.94 ns	89.40 ns	1.09	0.07	False

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
python3 .\performance\scripts\benchmarks_ci.py -f net8.0 --filter 'System.Text.Json.Tests.Utf8JsonReaderCommentsTests*'

System.Text.Json.Tests.Utf8JsonReaderCommentsTests.Utf8JsonReaderCommentParsing(CommentHandling: Skip, SegmentSize: 0, TestCase: ShortSingleLine)

ETL Files

Histogram

JIT Disasms

System.Text.Json.Tests.Utf8JsonReaderCommentsTests.Utf8JsonReaderCommentParsing(CommentHandling: Allow, SegmentSize: 0, TestCase: ShortSingleLine)

ETL Files

Histogram

JIT Disasms

System.Text.Json.Tests.Utf8JsonReaderCommentsTests.Utf8JsonReaderCommentParsing(CommentHandling: Allow, SegmentSize: 100, TestCase: ShortMultiLine)

ETL Files

Histogram

JIT Disasms

System.Text.Json.Tests.Utf8JsonReaderCommentsTests.Utf8JsonReaderCommentParsing(CommentHandling: Skip, SegmentSize: 0, TestCase: ShortMultiLine)

ETL Files

Histogram

JIT Disasms

Docs

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

[dotnet-policy-service]

Tagging subscribers to this area: @dotnet/area-system-text-json, @gregsdennis
See info in area-owners.md if you want to be subscribed.

[DrewScoggins]

Looks related to #101761

[EgorBo]

The problem here is that we have a very large struct System.Text.Json.Utf8JsonReader size=192 with 9 gc pointers, so 120 bytes are non-gc

[EgorBo]

A big chunk of time is spent inside the explicit memory barrier here.

[EgorBo]

cc @jkotas @MichalStrehovsky so looks like unconditional full memory barrier (no-op on x64) might make the bulk move slower than individual helpers (where dmb ish migth be skipped, e.g. in case if the dest address is not in the gc heap) - it eats 15% of perf numbers

I found a few tricks to speed up the copy algorithm, just not sure that will be enough.

Maybe we should adjust the logic that decides when to emit the bulk helper, e.g. to:

if (gcPointers >= 4 && gcSize >= fullSize/2)
    // insert the barrier