Re: [PATCH v6 00/16] zswap IAA compress batching

[Eric Biggers <ebiggers@kernel.org>]

On Wed, Feb 05, 2025 at 11:20:46PM -0800, Kanchana P Sridhar wrote:
> IAA Compression Batching:
> =========================
> 
> This patch-series introduces the use of the Intel Analytics Accelerator
> (IAA) for parallel batch compression of pages in large folios to improve
> zswap swapout latency.

So, zswap is passed a large folio to swap out, and it divides it into 4K pages
and compresses each independently.  The performance improvement in this patchset
comes entirely from compressing the folio's pages in parallel, synchronously,
using IAA.

Before even considering IAA and going through all the pain of supporting
batching with an off-CPU offload, wouldn't it make a lot more sense to try just
compressing each folio in software as a single unit?  Compared to the existing
approach of compressing the folio in 4K chunks, that should be much faster and
produce a much better compression ratio.  Compression algorithms are very much
designed for larger amounts of data, so that they can find more matches.

It looks like the mm subsystem used to always break up folios when swapping them
out, but that is now been fixed.  It looks like zswap just hasn't been updated
to do otherwise yet?

FWIW, here are some speed and compression ratio results I collected in a
compression benchmark module that tests feeding vmlinux (uncompressed_size:
26624 KiB) though zstd in 4 KiB page or 2 MiB folio-sized chunks:

zstd level 3, 4K chunks: 86 ms; compressed_size 9429 KiB
zstd level 3, 2M chunks: 57 ms; compressed_size 8251 KiB
zstd level 1, 4K chunks: 65 ms; compressed_size 9806 KiB
zstd level 1, 2M chunks: 34 ms; compressed_size 8878 KiB

The current zswap parameterization is "zstd level 3, 4K chunks".  I would
recommend "zstd level 1, 2M chunks", which would be 2.5 times as fast and give a
6% better compression ratio.

What is preventing zswap from compressing whole folios?

- Eric