Re: [PATCH v2 1/2] sched/numa: Drive NUMA task tick from execution context

[Hui Su]

On Tue, Sep 08, 2026 at 09:01:16AM -0700, Tim Chen wrote:
> I would vote for the real-time comparison. What do you think, Tim?
>
> Agree, I think keeping the real-time comparison makes sense.

Thanks for confirming, Tim.

I agree that the task-clock comparison is the better direction here. I have
incorporated it into the v4 series as patch 5/5, using the donor's
exec_start and a selection-time core_sched_start baseline instead of a
vruntime snapshot. The v4 series has now been sent.

This keeps the comparison in the task-clock domain and avoids mixing
vruntime deltas accumulated under an old weight with a slice calculated from
the current weight. The non-proxy predicate matched the existing
sum_exec_runtime predicate across the tested HZ and nice configurations, and
the proxy force-idle reproducer showed the donor task-clock delta advancing
while the donor sum_exec_runtime delta remained zero.

The patch is available here:

https://lore.kernel.org/r/20260909092901.2989564-6-sh_def@xxxxxxx

The v4 cover letter is available here:

https://lore.kernel.org/r/20260909092901.2989564-1-sh_def@xxxxxxx

Thanks,
Hui