Re: [PATCH v1 0/8] Batched zswap_store() with compression batching

[Kanchana P. Sridhar]

On Fri, Oct 16, 2026 at 2:48 PM Yosry Ahmed <yosry@xxxxxxxxxx> wrote:
>
> On Fri, Oct 16, 2026 at 12:05 PM Kanchana P. Sridhar
> <kanchanapsridhar2026@xxxxxxxxx> wrote:
> >
> > On Fri, Oct 16, 2026 at 9:14 AM Yosry Ahmed <yosry@xxxxxxxxxx> wrote:
> > >
> > > On Thu, Oct 8, 2026 at 11:28 AM Kanchana P. Sridhar
> > > <kanchanapsridhar2026@xxxxxxxxx> wrote:
> > > >
> > > > v1: Batched zswap_store() with compression batching
> > > > ===================================================
> > > >
> > > > This series is a spin-off of "v14: zswap compression batching with optimized
> > > > iaa_crypto driver" [1], and focuses on the generic crypto_acomp and zswap
> > > > compression batching enablers. Resuming work on this series following a gap
> > > > since v14 (January 2026). Please note that my email address and affiliation have
> > > > changed from kanchana.p.sridhar@xxxxxxxxx to kanchanapsridhar2026@xxxxxxxxx. I
> > > > have retained the original Intel sign-offs on prior patches for historical
> > > > attribution while signing off v1 of this series with my current address.
> > > >
> > > > To summarize the main contribution of this patch series: Batching is introduced
> > > > in zswap_store() of large folios: we carve out a large folio into batches of
> > > > ZSWAP_MAX_BATCH_SIZE (8 pages) for software compressors, and store each
> > > > batch. Moving the fundamental operations that need to be happen to store a page
> > > > to now work at the granularity of a batch offers opportunities to optimize them
> > > > from these perspectives:
> > > >
> > > > 1) Utilize existing kernel API that minimize obtaining locks for bulk
> > > >    operations, e.g. kmem_cache_alloc_bulk()/kmem_cache_free_bulk().
> > > >
> > > > 2) We keep the working set for each of the core store functionalities (alloc
> > > >    entries, compress pages, store pages in zsmalloc) to be focused on a
> > > >    batch, thereby minimizing cache evictions and benefiting from cache
> > > >    locality.
> > > >
> > > > 3) Along the lines of (1), the main zswap_compress() batching routine obtains
> > > >    the acomp_ctx mutex lock once per batch, and releases it after it has
> > > >    processed all pages in that batch.
> > > >
> > > > 4) Any zswap store functionality that works on a batch can potentially be
> > > >    parallelized as a next step, to further improve performance.
> > > >
> > > > Performance data: usemem with 60% over-commit:
> > > > ==============================================
> > > > Validation was done with mm-unstable as of 10-3-2026 (Baseline) and with this
> > > > patch series.
> > > >
> > > > Testing was done on a Chromebook with a QEMU "vng" VM running in Crostini.
> > > >
> > > > Chromebook Diagnostics reports:
> > > >   Intel(R) Core(TM) Ultra 5 115U (10 CPUs, 4.200GHz)
> > > >   RAM: 16GB
> > > >
> > > > vng setup:
> > > >          Memory: 7G
> > > >            Swap: 3G
> > > >     Pinned CPUs: 4 (E-Cores: 4,5,6,7)
> > > >
> > > > zswap compressor:
> > > >   ZSTD
> > > >
> > > > mTHP enabled:
> > > >   64kB
> > > >
> > > > vm-scalability/usemem setup:
> > > >
> > > >   cgroup memory.high: 3G
> > > >   4 usemem processes, each allocates 1200M anonymous memory, reads the
> > > >   Silesia dataset [2] from a pre-fork-allocated buffer into the 1200M,
> > > >   writes to each unsigned long in the 1200M, sleeps for 10 sec, then
> > > >   frees the memory and exits.
> > > >
> > > >     usemem --read-silesia -w -O -b 1 -s 10 -n 4 1200M
> > >
> > > What is --read-silesia? Is this a local change you made to usemem?
> >
> > Yes, this is a local change I made because I wanted to get compressed
> > lengths besides 565, which is the only compressed length of data
> > swapped out from the original usemem.
> >
> > To replicate my setup:
>
> Thanks for sharing the details. I am having trouble reproducing the
> results, mainly because the standard deviation seems to be too high to
> make conclusions. What is the standard deviation in your runs?

Sure thing! Let me know if it would be useful to upstream the
--read-silesia option to the vm-scalability tree.

Yes, standard deviation is considerable, which is why I run usemem
consecutively for 10 iterations and report the mean Average
throughput. I could also report the median going forward if that
helps. The cover letter vmstats are after 10 iterations. The last set
of data I reported in this thread after applying patch 8 [batched
version of zswap_compress()], had a mean of 20,946 KB/s, standard
deviation of 5,391 KB/s and median of 22,477 KB/s. For reference, the
stddev for the baseline mm-unstable was 5,397 KB/s.

I have found this methodology to be reliable in drawing conclusions
because as zswap_store() performance improves, the outliers are fewer
and the data is more clustered near the mean.

>
> Also, I tried measuring the latency of zswap_store() itself through
> bpftrace, and it doesn't seem to change much throughout the series.

I ran bpftrace today to gather zswap_compress() latency per-page,
before and after patch 8. It shows a clear performance improvement
with compress batching for the reasons mentioned before:

Sequential_zswap_compress  nr_samples     p25     p50      p75
p95       p99
=================================================================================
zswap_compress_ns_per_page  1,722,489  43,947  57,791   88,489
130,165   323,961

Batched_zswap_compress     nr_samples     p25     p50      p75
p95       p99
=================================================================================
zswap_compress_ns_per_page    339,088  42,081  52,633   63,185
118,461   191,710
=================================================================================
 Latency saving (ns)                   -1,866  -5,159  -25,305
-11,704  -132,251
=================================================================================

Hope this helps. Thanks!