GH-38452: [C++][Benchmark] Adding benchmark for LZ4/Snappy Compression

[mapleFU]

Rationale for this change

This patch add LZ4 (LZ4_RAW in Parquet standard) and Snappy compression/decompression benchmark.

What changes are included in this PR?

Add groups of benchmarks.

Are these changes tested?

no

Are there any user-facing changes?

no

• Closes: [C++][Benchmark] Adding benchmark for LZ4/Snappy Compression #38452

[github-actions]

⚠️ GitHub issue #38452 has been automatically assigned in GitHub to PR creator.

[pitrou]

I get the following error with this PR:

/home/antoine/arrow/dev/cpp/src/arrow/result.cc:28: ValueOrDie called on an error: NotImplemented: Streaming compression unsupported with LZ4 raw format. Try using LZ4 frame format instead.

[mapleFU]

This is my fault, I just add the code without checking.

I've re-check the logic here. The LZ4 and Snappy doesn't support Streaming ( probabily this is the reason origin tests doesn't add them). Now I've remove the streaming for these two Codec.

[pitrou]

For the record, benchmark numbers here:

ReferenceStreamingCompression<Compression::GZIP>         170571860 ns    169802391 ns            4 bytes_per_second=47.1136M/s ratio=6.95102
ReferenceCompression<Compression::GZIP>                  171569422 ns    171399551 ns            4 bytes_per_second=46.6746M/s ratio=6.95102
ReferenceStreamingDecompression<Compression::GZIP>        12079114 ns     12075244 ns           58 bytes_per_second=662.513M/s ratio=6.95102
ReferenceDecompression<Compression::GZIP>                 11983466 ns     11926139 ns           58 bytes_per_second=670.795M/s ratio=6.95102

ReferenceStreamingCompression<Compression::BROTLI>       254898609 ns    253547102 ns            3 bytes_per_second=31.5523M/s ratio=8.31174
ReferenceCompression<Compression::BROTLI>                258946902 ns    257261268 ns            3 bytes_per_second=31.0968M/s ratio=8.31175
ReferenceStreamingDecompression<Compression::BROTLI>       8415786 ns      8343701 ns           83 bytes_per_second=958.807M/s ratio=8.31174
ReferenceDecompression<Compression::BROTLI>                7481414 ns      7443618 ns           94 bytes_per_second=1074.75M/s ratio=8.31175

ReferenceStreamingCompression<Compression::ZSTD>          17918513 ns     17833898 ns           39 bytes_per_second=448.584M/s ratio=6.876
ReferenceCompression<Compression::ZSTD>                   16608187 ns     16545738 ns           42 bytes_per_second=483.508M/s ratio=6.8771
ReferenceStreamingDecompression<Compression::ZSTD>         5673372 ns      5646730 ns          123 bytes_per_second=1.38354G/s ratio=6.876
ReferenceDecompression<Compression::ZSTD>                  5227017 ns      5202190 ns          132 bytes_per_second=1.50177G/s ratio=6.8771

ReferenceStreamingCompression<Compression::LZ4_FRAME>     14577950 ns     14570210 ns           47 bytes_per_second=549.066M/s ratio=3.52824
ReferenceCompression<Compression::LZ4_FRAME>              12065500 ns     12007357 ns           58 bytes_per_second=666.258M/s ratio=3.52824
ReferenceStreamingDecompression<Compression::LZ4_FRAME>    2008312 ns      1998786 ns          349 bytes_per_second=3.90862G/s ratio=3.52824
ReferenceDecompression<Compression::LZ4_FRAME>             1983992 ns      1972508 ns          355 bytes_per_second=3.96069G/s ratio=3.52824

ReferenceCompression<Compression::LZ4>                    11887179 ns     11835535 ns           59 bytes_per_second=675.931M/s ratio=3.53112
ReferenceDecompression<Compression::LZ4>                   1935571 ns      1924247 ns          363 bytes_per_second=4.06003G/s ratio=3.53112

ReferenceCompression<Compression::SNAPPY>                 11624693 ns     11564830 ns           61 bytes_per_second=691.752M/s ratio=3.58312
ReferenceDecompression<Compression::SNAPPY>                7049123 ns      7016010 ns          100 bytes_per_second=1.11352G/s ratio=3.58312

[wgtmac]

For the record, benchmark numbers here:

ReferenceStreamingCompression<Compression::GZIP>         170571860 ns    169802391 ns            4 bytes_per_second=47.1136M/s ratio=6.95102
ReferenceCompression<Compression::GZIP>                  171569422 ns    171399551 ns            4 bytes_per_second=46.6746M/s ratio=6.95102
ReferenceStreamingDecompression<Compression::GZIP>        12079114 ns     12075244 ns           58 bytes_per_second=662.513M/s ratio=6.95102
ReferenceDecompression<Compression::GZIP>                 11983466 ns     11926139 ns           58 bytes_per_second=670.795M/s ratio=6.95102

ReferenceStreamingCompression<Compression::BROTLI>       254898609 ns    253547102 ns            3 bytes_per_second=31.5523M/s ratio=8.31174
ReferenceCompression<Compression::BROTLI>                258946902 ns    257261268 ns            3 bytes_per_second=31.0968M/s ratio=8.31175
ReferenceStreamingDecompression<Compression::BROTLI>       8415786 ns      8343701 ns           83 bytes_per_second=958.807M/s ratio=8.31174
ReferenceDecompression<Compression::BROTLI>                7481414 ns      7443618 ns           94 bytes_per_second=1074.75M/s ratio=8.31175

ReferenceStreamingCompression<Compression::ZSTD>          17918513 ns     17833898 ns           39 bytes_per_second=448.584M/s ratio=6.876
ReferenceCompression<Compression::ZSTD>                   16608187 ns     16545738 ns           42 bytes_per_second=483.508M/s ratio=6.8771
ReferenceStreamingDecompression<Compression::ZSTD>         5673372 ns      5646730 ns          123 bytes_per_second=1.38354G/s ratio=6.876
ReferenceDecompression<Compression::ZSTD>                  5227017 ns      5202190 ns          132 bytes_per_second=1.50177G/s ratio=6.8771

ReferenceStreamingCompression<Compression::LZ4_FRAME>     14577950 ns     14570210 ns           47 bytes_per_second=549.066M/s ratio=3.52824
ReferenceCompression<Compression::LZ4_FRAME>              12065500 ns     12007357 ns           58 bytes_per_second=666.258M/s ratio=3.52824
ReferenceStreamingDecompression<Compression::LZ4_FRAME>    2008312 ns      1998786 ns          349 bytes_per_second=3.90862G/s ratio=3.52824
ReferenceDecompression<Compression::LZ4_FRAME>             1983992 ns      1972508 ns          355 bytes_per_second=3.96069G/s ratio=3.52824

ReferenceCompression<Compression::LZ4>                    11887179 ns     11835535 ns           59 bytes_per_second=675.931M/s ratio=3.53112
ReferenceDecompression<Compression::LZ4>                   1935571 ns      1924247 ns          363 bytes_per_second=4.06003G/s ratio=3.53112

ReferenceCompression<Compression::SNAPPY>                 11624693 ns     11564830 ns           61 bytes_per_second=691.752M/s ratio=3.58312
ReferenceDecompression<Compression::SNAPPY>                7049123 ns      7016010 ns          100 bytes_per_second=1.11352G/s ratio=3.58312

It seems that streaming (de)compression is generally slower than the non-streaming parity.

[pitrou]

There may be some memory allocation costs for streaming (de)compression, while non-streaming is stateless.

[conbench-apache-arrow]

After merging your PR, Conbench analyzed the 6 benchmarking runs that have been run so far on merge-commit 73589dd.

There were no benchmark performance regressions. 🎉

The full Conbench report has more details. It also includes information about 3 possible false positives for unstable benchmarks that are known to sometimes produce them.