Re: [PATCH] sched/fair: avoid creating misfits during cache-aware balancing

[Tim Chen <tim.c.chen@linux.intel.com>]

On Mon, 2026-09-14 at 10:59 +0800, KobaK wrote:
> Hi Tim,
> 
> Thanks for the cache-aware load-balancing misfit patch. The tradeoff
> you describe (LLC locality vs capacity loss) is clear.
> 
> I have been trying to understand how this showed up in testing. The
> cover explains the failure mode, but I did not find a note on the
> observation method.
> 
> Would you mind sharing how you tested and observed it? In particular:
> which hybrid generation, what workload, and whether you used
> schedstat lb_imbalance_misfit or another signal.
> 
> Thanks,
> Koba

Kobak,

My colleague Ricardo was doing tests on the Nova Lake CPU which has
LP-E cores on one LLC domain and the E-cores/P-cores on
other LLC domain. He pinned a process with
threads on the LP-E cores and then unpin it.

His expectation was that the process should move to
the more capable cores but that did
not happen because cache aware scheduling took precedent.
So the patch was meant to fix this behavior.
This can be seen easily with mpstat.

Cc Ricardo here and he can provide more details if you
are interested.

Tim