Re: [PATCH v4] sched/fair: Introduce SIS_UTIL to search idle CPU based on sum of util_avg

[Mel Gorman]

On Mon, Jun 13, 2022 at 04:06:36PM +0800, Chen Yu wrote:
> On Mon, Jun 13, 2022 at 03:40:52PM +0800, Yicong Yang wrote:
> > On 2022/6/13 0:34, Chen Yu wrote:
> > >  
> [cut...]
> > >  #define NUMA_IMBALANCE_MIN 2
> > > diff --git a/kernel/sched/features.h b/kernel/sched/features.h
> > > index 1cf435bbcd9c..3334a1b93fc6 100644
> > > --- a/kernel/sched/features.h
> > > +++ b/kernel/sched/features.h
> > > @@ -61,6 +61,7 @@ SCHED_FEAT(TTWU_QUEUE, true)
> > >   * When doing wakeups, attempt to limit superfluous scans of the LLC domain.
> > >   */
> > >  SCHED_FEAT(SIS_PROP, true)
> > > +SCHED_FEAT(SIS_UTIL, true)
> > >  
> > 
> > confused here that shouldn't we have SCHED_FEAT(SIS_PROP, false)? With SIS_UTIL enabled, SIS_PROP will have no
> > effect since nr is overridden by SIS_UTIL.
> Yes, no matter what SIS_PROP is set, the result of SIS_UTIL would be used to decide
> the scan depth. We don't change the default value of SIS_PROP here, as this patch
> tends to only touch one feature at one time. And the options could be tuned by user via
> sysfs manually. Besides, the target is to replace SIS_PROP with another search policy,
> Peter mentioned that "And ideally we're remove SIS_PROP after a few releases if this
> works out", so I assume that changing the default value of SIS_PROP does not matter
> in current patch.
> 

I had expected it to be disabled given that SIS_PROP does work to
calculcate nr, then discards it, and uses SIS_UTIL. If SIS_UTIL shows a
regression and reports a bug, the first step would be to disable
SIS_UTIL and enable SIS_PROP via sched_feat.

-- 
Mel Gorman
SUSE Labs