Re: [linus:master] [mm] f1a7941243: unixbench.score -19.2% regression

On Sun, Feb 26, 2023 at 10:35 PM Yin, Fengwei [off-list ref] wrote:

Hi Andrew, Shakeel,

On Tue, 2023-01-31 at 10:26 -0800, Shakeel Butt wrote:

quoted

+per-cpu memory maintainers for FYI.

Thread started at
https://lore.kernel.org/linux-mm/202301301057.e55dad5b-oliver.sang@intel.com/ (local)

On Mon, Jan 30, 2023 at 9:57 PM Shakeel Butt [off-list ref]
wrote:

quoted

[...]

quoted

quoted

We could cut down the number of calls to pcpu_alloc() by a factor
of 4
by having a pcpu_alloc_bulk() that would allocate all four RSS
counters
at once.

Just throwing out ideas ...

Thanks, I will take a stab at pcpu_alloc_bulk() and will share the
result tomorrow.

OK, not a one day effort.

Andrew, which option do you prefer?

1. Keep the patches as the test workload (fork ping pong) is not a
representative of real world workload and work on improving
pcpu_alloc() for 6.4+.

2. Revert the patches for now, improve pcpu_alloc() and re-introduce
the patches once we confirm that fork-ping-pong is not regressed
anymore.

This performance regression still can be reproduced on latest master
branch. So we took option1 here. Right? Thanks.

Yes unless some real workload reports regression.

Shakeel