Re: [PATCH v3 7/7] sched/eevdf: Move to a single runqueue

[Chengfeng Lin]

Hi Mike,

Thanks for clarifying the setup. I understand that you see the effect
when pinned wakelat and eight pinned CPU-bound workers compete within
the same cgroup.

If you make further changes in this area, I'd be happy to rerun the
context-switch benchmark across the full range of worker counts and
share the results.

Thanks,
Chengfeng



On Sun, 01 Nov 2026 18:46:43 +0100, Mike Galbraith <efault@xxxxxx> wrote:
> On Sun, 2026-11-01 at 22:37 +0800, Chengfeng Lin wrote:
> >
> > Is this the path you had in mind, or were you referring to a different
> > update? I'd be happy to test another candidate change.
>
> Where I see negative impact of the __dequeue_task() update is when
> probe components compete in one cgroup, ie pinned wakelat and 8 pinned
> hogs used to measure the delta between smp and concur CPU distribution
> modes during cgroup vs cgroup testing, all stuffed into one cgroup.
>
> Do whatever you like there, just keep update_curr() far away from
> cfs_rq_of(se) and probe goes silent.  For virgin flat in isolation,
> that was the only signal the probe generated.
>
>       -Mike