Re: [PATCH v2] sched/fair: fix 1 task per CPU

[Peter Zijlstra]

On Thu, Dec 13, 2018 at 12:04:20PM +0100, Vincent Guittot wrote:
> On Thu, 13 Dec 2018 at 11:44, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Mon, Sep 10, 2018 at 04:43:09PM +0200, Vincent Guittot wrote:
> > > When CPUs have different capacity because of RT/DL tasks or
> > > micro-architecture or max frequency differences, there are situation where
> > > the imbalance is not correctly set to migrate waiting task on the idle CPU.
> > >
> > > The UC uses the force_balance case:
> > >
> > >       if (env->idle != CPU_NOT_IDLE && group_has_capacity(env, local) &&
> > >           busiest->group_no_capacity)
> > >               goto force_balance;
> > >
> > > But calculate_imbalance fails to set the right amount of load to migrate
> > > a task because of the special condition:
> > >
> > >   busiest->avg_load <= sds->avg_load || local->avg_load >= sds->avg_load)
> > >
> > > Add in fix_small_imbalance, this special case that triggered the force
> > > balance in order to make sure that the amount of load to migrate will be
> > > enough.
> >
> > So I think this patch is going in the wrong direction for a number of
> > reasons:
> >
> >  - we'd like to get rid of fix_small_imbalance(), and this adds to it;
> >
> >  - the whole load_per_task stuff is terminally broken, it _cannot_ work
> >    right.
> >
> >
> > What I've suggested in the past is parameterizing the load balancer and
> > picking different criteria to balance on:
> 
> This patch is clearly a fix of the current implementation.
> What you suggest below makes sense but implies a significant rework in
> the calculate_imbalance and the load_balancer in general and will need
> more time to reach a stable state.
> Nevertheless, I will have a look at that
> 
> I imagine that your feedback for https://lkml.org/lkml/2018/10/2/283
> will be the same ?

No; those actually look ok. It is mostly that I really don't think
load_per_task makes any kind of sense.

It sorta works when all tasks are of the same weight, but if you start
using nice -- or way worse, cgroups -- then the number is complete
bollocks.