Re: [PATCH] sched_ext: introduce LLC awareness to the default idle selection policy

[Andrea Righi <arighi@nvidia.com>]

On Mon, Oct 21, 2024 at 10:01:27AM -1000, Tejun Heo wrote:
> External email: Use caution opening links or attachments
> 
> 
> Hello,
> 
> Overall, I think this is a great idea.
> 
> On Mon, Oct 21, 2024 at 09:13:04AM +0200, Andrea Righi wrote:
> ...
> > +     /*
> > +      * Determine the task's LLC domain.
> > +      */
> > +     sd = rcu_dereference(per_cpu(sd_llc, prev_cpu));
> > +     if (sd)
> > +             cpumask_and(llc_cpus, sched_domain_span(sd), p->cpus_ptr);
> > +     else
> > +             cpumask_copy(llc_cpus, p->cpus_ptr);
> 
> However, I wonder whether we can be a bit more efficient here. Always
> copying cpumasks can become noticeable in larger machines. It should be
> possible to cover most common cases without copying cpumasks - e.g. tasks
> which don't have any cpumask restrictions or affine within a single LLC
> (including tasks restricted to one CPU) don't need to compute a new cpumask
> each time. They can use either sched_domain_span() or p->cpus_ptr directly.

I agree, I was also thinking to improve this part to avoid doing the
copy. And I have a few other changes to apply, I'll send a v2 soon.

Thanks!
-Andrea

> 
> Thanks.
> 
> --
> tejun