Re: [PATCH 4/7] sched/fair: Avoid unnecessary balancing of asymmetric capacity groups
From: Peter Zijlstra
Date: Tue Feb 20 2018 - 13:26:22 EST
On Tue, Feb 20, 2018 at 04:33:52PM +0000, Morten Rasmussen wrote:
> On Mon, Feb 19, 2018 at 04:10:11PM +0100, Peter Zijlstra wrote:
> > On Thu, Feb 15, 2018 at 04:20:51PM +0000, Morten Rasmussen wrote:
> > > +/*
> > > + * group_similar_cpu_capacity: Returns true if the minimum capacity of the
> > > + * compared groups differ by less than 12.5%.
> > > + */
> > > +static inline bool
> > > +group_similar_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
> > > +{
> > > + long diff = sg->sgc->min_capacity - ref->sgc->min_capacity;
> > > + long max = max(sg->sgc->min_capacity, ref->sgc->min_capacity);
> > > +
> > > + return abs(diff) < max >> 3;
> > > +}
> >
> > This seems a fairly random and dodgy heuristic.
>
> I can't deny that :-)
>
> We need to somehow figure out if we are doing asymmetric cpu capacity
> balancing or normal SMP balancing. We probably don't care about
> migrating tasks if the capacities are nearly identical. But how much is
> 'nearly'?
>
> We could make it strictly equal as long as sgc->min_capacity is based on
> capacity_orig. If we let things like rt-pressure influence
> sgc->min_capacity, it might become a mess.
See, that is the problem, I think it this min_capacity thing is
influenced by rt-pressure and the like.
See update_cpu_capacity(), min_capacity is set after we add the RT scale
factor thingy, and then update_group_capacity() filters the min of the
whole group. The thing only ever goes down.
But this means that if a big CPU has a very high IRQ/RT load, its
capacity will dip below that of a little core and min_capacity for the
big group as a whole will appear smaller than that of the little group.
Or am I now terminally confused again?