Explore using PyBytesWriter API for compression libraries output buffers

[emmatyping]

Feature or enhancement

The new PyBytesWriter() API is fast and easy to use. I expect it will bring a nice improvement both to maintainability and speed for compression output buffer management.

I have some perf recordings showing that a large portion (>50%!) of time in decompression for a mix of data sizes (1K, 1M, 1G) is in _BlocksOutputBuffer_Finish, re-assembling the output buffer.

I also made a very hacky modification to pycore_blocks_output_buffer.h to use PyBytesWriter() and found it greatly sped up decompression time:

The below two tests are operating on compressed enwiki content with zstd compression.

test	main	PyBytesWriter()
decompress 1M	2.15ms	1.65ms
decompress 1G	2.2s	1.73s

Those are 25-30% speedups!

I think this is enough to motivate a refactor of this code to use PyBytesWriter() and benchmark against the current implementation across compression modules and data sizes.

cc @vstinner for viz

Linked PRs

• gh-139877: Use PyBytesWriter in pycore_blocks_output_buffer.h #139976

[cmaloney]

_BlocksOutputBuffer_Finish seems to be doing a "".join() on self->list. "stringlib" has a fast/fairly tuned implementation of that, wondering if the end join could just become something like:

    PyObject *new_buffer = PyObject_CallMethodOneArg(
        Py_GetConstant(Py_CONSTANT_EMPTY_BYTES), &_Py_ID(join), self->list);

[emmatyping]

Good idea, I will also test that. I expect PyBytesWriter will still be faster because of locality, but I can benchmark both.

[emmatyping]

I have more data, looks like not all of the (de)compressors rely on the output buffer code as much as zstd. That makes sense since the actual compression code is slower so the output buffer allocations are relatively less important.

I've started with zlib as that is the most used format most likely. It's also the (de)compressor for wheels so performance gains there have huge multipliers :)

My system:
AMD Ryzen 9 9950X
6000 MT/s DDR5 RAM
Running Fedora 42

Here's main vs using "".join():

zlib.compress(1M): Mean +- std dev: [main] 13.5 ms +- 0.1 ms -> [bytes_join] 13.5 ms +- 0.1 ms: 1.00x slower
zlib.compress(1G): Mean +- std dev: [main] 11.4 sec +- 0.0 sec -> [bytes_join] 11.4 sec +- 0.0 sec: 1.00x slower
zlib.decompress(1K): Mean +- std dev: [main] 1.42 us +- 0.01 us -> [bytes_join] 1.44 us +- 0.01 us: 1.01x slower
zlib.decompress(1M): Mean +- std dev: [main] 1.29 ms +- 0.00 ms -> [bytes_join] 1.20 ms +- 0.00 ms: 1.07x faster
zlib.decompress(1G): Mean +- std dev: [main] 1.36 sec +- 0.00 sec -> [bytes_join] 1.39 sec +- 0.01 sec: 1.02x slower

Benchmark hidden because not significant (1): zlib.compress(1K)
Geometric mean: 1.00x faster

Note that I had to add a _PyBytes_Resize immediately after the join() because the output bytes object needs to be truncated to the length of the output data.

Here's comparing main vs PyBytesWriter:

zlib.compress(1M): Mean +- std dev: [main] 13.5 ms +- 0.1 ms -> [pybyteswriter] 13.4 ms +- 0.0 ms: 1.00x faster
zlib.compress(1G): Mean +- std dev: [main] 11.4 sec +- 0.0 sec -> [pybyteswriter] 11.3 sec +- 0.0 sec: 1.00x faster
zlib.decompress(1K): Mean +- std dev: [main] 1.42 us +- 0.01 us -> [pybyteswriter] 1.39 us +- 0.01 us: 1.02x faster
zlib.decompress(1M): Mean +- std dev: [main] 1.29 ms +- 0.00 ms -> [pybyteswriter] 1.17 ms +- 0.00 ms: 1.10x faster
zlib.decompress(1G): Mean +- std dev: [main] 1.36 sec +- 0.00 sec -> [pybyteswriter] 1.17 sec +- 0.00 sec: 1.17x faster

Benchmark hidden because not significant (1): zlib.compress(1K)

Geometric mean: 1.05x faster

5% overall is not bad! 10% faster decompression for data >1MB is also great, that will be a big win for wheels and many other use cases I imagine.

I assume the numbers for xz and bz2 are pretty similar or a more minor win, since those (de)compressors are as slow or slower than zlib.

I'm going to collect numbers for zstd and verify my initial numbers next, then work on a patch.

[emmatyping]

Okay! As expected, the improvements are a lot better for zstd (using default compression level 3):

zstd.compress(1K): Mean +- std dev: [main_zstd_3] 3.01 us +- 0.03 us -> [pybyteswriter_zstd_3] 3.00 us +- 0.03 us: 1.01x faster
zstd.compress(1M): Mean +- std dev: [main_zstd_3] 2.92 ms +- 0.02 ms -> [pybyteswriter_zstd_3] 2.89 ms +- 0.02 ms: 1.01x faster
zstd.compress(1G): Mean +- std dev: [main_zstd_3] 2.72 sec +- 0.01 sec -> [pybyteswriter_zstd_3] 2.67 sec +- 0.01 sec: 1.02x faster
zstd.decompress(1K): Mean +- std dev: [main_zstd_3] 1.40 us +- 0.01 us -> [pybyteswriter_zstd_3] 1.38 us +- 0.01 us: 1.01x faster
zstd.decompress(1M): Mean +- std dev: [main_zstd_3] 734 us +- 4 us -> [pybyteswriter_zstd_3] 546 us +- 3 us: 1.34x faster
zstd.decompress(1G): Mean +- std dev: [main_zstd_3] 790 ms +- 4 ms -> [pybyteswriter_zstd_3] 634 ms +- 3 ms: 1.25x faster

Geometric mean: 1.10x faster