Re: [PATCH RFC 0/3] zram: avoid copying spanning LZ4 objects before decompression
From: Xueyuan Chen
Date: Sat Nov 21 2026 - 20:59:15 EST
On 2026-11-21 12:15 -0800, Eric Biggers wrote:
> Hi Xueyuan,
>
> On Sun, Nov 22, 2026 at 01:09:23AM +0800, Xueyuan Chen wrote:
> > I measured the time to swap in 1 GiB of synthetic data on arm64 with
> > 4 KiB pages. Each page started with pseudorandom data of one of 16
> > lengths, followed by zeros. On this workload, 32.6% of compressed
> > reads spanned two pages with LZ4 and 34.4% with LZ4HC.
> >
> > The table shows median swap-in times over seven pairs. The last two
> > columns summarize the percentage change within each pair.
> >
> > Codec Workers Baseline (ms) Patched (ms) Median Range
> > LZ4 1 1403.81 1344.45 -4.21% -4.64% to -1.61%
> > LZ4 8 222.57 216.36 -2.73% -4.19% to -1.97%
> > LZ4HC 1 1403.35 1336.99 -4.74% -5.07% to -2.90%
> > LZ4HC 8 222.10 215.77 -3.16% -4.39% to -1.27%
> >
> > These timings include both skipping the copy and switching decoders.
> >
> > I used Codex to help implement this series.
> > Do you think the swap-in improvement justifies maintaining a second
> > LZ4 decoder?
> >
> > Assisted-by: LLM
> > Signed-off-by: Xueyuan Chen <xueyuan.chen21@xxxxxxxxx>
>
> LZ4_decompress_safe_split() isn't actually very well optimized. While
> it avoids the copy of the source data into a contiguous buffer, it
> copies every match and literal run with a variable-length memcpy(),
> which on most data is a very slow way to implement LZ decompression.
>
> I suspect it looked good on your synthetic benchmark (pseudorandom data
> followed by 0's) only because that tended to generate 1 very long
> literal run and 1 very long match. Whereas normal data would have lots
> of interleaved literal runs and matches of much shorter lengths where an
> external memcpy() call does very poorly; that is why the existing LZ4
> decompression code does not do it that way.
>
> To some extent the literal run and match copy performance is fixable
> with a better implementation. However, having to handle two different
> source buffers still introduces overhead that cannot go away.
>
> So I feel this approach is probably not worth it. If you'd like to
> continue to explore it, it would need to be tested on different data and
> would need to be properly optimized.
>
> - Eric
>
>
Hi Eric,
Thanks for taking a look. I was also unsure how representative the
synthetic workload was, so I was hoping to get some feedback from
experienced kernel developers.
I'll collect 1 GiB of data from zram on an Android phone and rerun the
swap-in benchmark with it.