On Fri, Jun 9, 2023 at 7:04 AM Marc Zyngier [off-list ref] wrote:

On Fri, 09 Jun 2023 01:59:35 +0100,
Yu Zhao [off-list ref] wrote:

quoted

TLDR
====
Apache Spark spent 12% less time sorting four billion random integers twenty times (in ~4 hours) after this patchset [1].

Why are the 3 architectures you have considered being evaluated with 3
different benchmarks?

I was hoping people having special interests in different archs might
try to reproduce the benchmarks that I didn't report (but did cover)
and see what happens.

I am not suspecting you to have cherry-picked
the best results

I'm generally very conservative when reporting *synthetic* results.
For example, the same memcached benchmark used on powerpc yielded >50%
improvement on aarch64, because the default Ubuntu Kconfig uses 64KB
base page size for powerpc but 4KB for aarch64. (Before the series,
the reclaim (swap) path takes kvm->mmu_lock for *write* on O(nr of all
pages to consider); after the series, it becomes O(actual nr of pages
to swap), which is <10% given how the benchmark was set up.)

          Ops/sec  Avg. Latency  p50 Latency  p99 Latency  p99.9 Latency
------------------------------------------------------------------------
Before  639511.40       0.09940      0.04700      0.27100       22.52700
After   974184.60       0.06471      0.04700      0.15900        3.75900

but I'd really like to see a variety of benchmarks
that exercise this stuff differently.

I'd be happy to try other synthetic workloads that people think that
are relatively representative. Also, I've backported the series and
started an A/B experiment involving ~1 million devices (real-world
workloads). We should have the preliminary results by the time I post
the next version.