Re: [RFC PATCH v8] sched: Fix performance regression introduced by mm_cid

[Mathieu Desnoyers]

On 2023-04-18 12:01, Mathieu Desnoyers wrote:
> On 2023-04-18 07:21, Aaron Lu wrote:
> > On Mon, Apr 17, 2023 at 11:08:31AM -0400, Mathieu Desnoyers wrote:
> > > Introduce per-mm/cpu current concurrency id (mm_cid) to fix a PostgreSQL
> > > sysbench regression reported by Aaron Lu.
> >
> > For postgres_sysbench on SPR:
> > sched_mm_cid_migrate_to() is in the range of 0.1x% - 0.4x%, mm_cid_get()
> > is in the range of 0.1x% - 0.3x%. Other cid functions are pretty minor.
> >
> > For hackbench on SPR:
> > ched_mm_cid_migrate_to() is about 3%-4%, mm_cid_get() is about 7%-8%,
> > other cid functions are pretty minor.
>
> It's a bit higher than I would have expected for hackbench.
>
> Can you run with the attached schedstats patch applied and boot
> with schedstats=enable ? Let me know how the counters behave please,
> because I cannot reproduce anything unexpected on my machine.

Also, just to confirm, do you still use this hackbench invocation ?

hackbench --pipe --threads -l 500000

Thanks,

Mathieu

> Thanks,
>
> Mathieu

--
Mathieu Desnoyers
EfficiOS Inc.
https://www.efficios.com