Performance regressions found in Brotli v1.1.0

[carlossanlop]

The dotnet/runtime repo recently updated its in-tree copy of its Brotli dependency from v1.0.9 to v1.1.0.

We ran our performance benchmarks, and we noticed regressions in Linux-x64 and Windows-x64.

Most of our regression results were under 15%, but 6 results were between 39% an 49%.

Here's a list of the tests that showed regressions:

Windows x64

• 49% - Decompress_WithState(level: Fastest, file: "TestDocument.pdf")
• 47% - Decompress_WithState(level: Optimal, file: "TestDocument.pdf")
• 47% - Decompress_WithoutState(level: Fastest, file: "TestDocument.pdf")
• 47% - Decompress_WithoutState(level: Optimal, file: "TestDocument.pdf")
• 39% - Decompress(level: Fastest, file: "TestDocument.pdf")
• 39% - Decompress(level: Optimal, file: "TestDocument.pdf")
• 13% - Decompress_WithoutState(level: Fastest, file: "alice29.txt")
• 12% - Decompress_WithoutState(level: Optimal, file: "alice29.txt")
• 12% - Decompress_WithState(level: Fastest, file: "alice29.txt")
• 11% - Decompress_WithState(level: Optimal, file: "alice29.txt")
• 11% - Decompress(level: Fastest, file: "alice29.txt")
• 10% - Decompress(level: Optimal: file: "alice29.txt")
• 10% - Decompress_WithState(level: Fastest, file: "sum")
• 10% - Decompress_WithoutState(level: Fastest, file: "sum")
• 09% - Decompress(level: Fastest, file: "sum")

Linux x64

• 16% - Compress_WithoutState(level: Optimal, file: "TestDocument.pdf")
• 16% - Compress_WithState(level: Optimal, file: "TestDocument.pdf")
• 15% - Compress_WithoutState(level: Optimal, file: "alice29.txt")
• 15% - Compress_WithoutState(level: Optimal, file: "sum")
• 14% - Compress_WithState(level: Optimal, file: "sum")
• 13% - Compress_WithState(level: Optimal, file: "alice29.txt")

Some more info:

• Our benchmark methods can be found [here](https://github.com/dotnet/performance/blob/main/src/benchmarks/micro/libraries/System.IO.Compression/Brotli.cs Files tested: https://github.com/dotnet/performance/tree/main/src/benchmarks/micro/libraries/System.IO.Compression/TestData).
• Our in-tree copy of Brotli v1.1.0 is found here.
• The managed code that consumes the in-tree copy of Brotli can be found here.
• These are the mappings of the CompressionLevel enum to Brotli quality values:
• "Fastest" => 1
• "Optimal" => 4:

Unfortunately, the v1.1.0 tag has no mention of known perf regressions, so we were wondering if you were aware of these results and if you could help us figure out the possible root causes.

[eustas]

Hello. Thanks for the report. I'll investigate soon.

We thoroughly test and polish changes until they are at least as performant as before. Likely regression happens on arch / compiler / OS combination we do not track.

[carlossanlop]

Likely regression happens on arch / compiler / OS combination we do not track.

We found them in Linux-x64 and Windows-x64.