Re: [PATCH] fix scheduler regression from "sched/fair: Rework load_balance()"

[Chris Mason]

On 26 Oct 2020, at 11:05, Chris Mason wrote:

> On 26 Oct 2020, at 10:24, Vincent Guittot wrote:
>
> > Could you try the fix below ?
> >
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -9049,7 +9049,8 @@ static inline void calculate_imbalance(struct 
> > lb_env *env, struct sd_lb_stats *s
> >          * emptying busiest.
> >          */
> >         if (local->group_type == group_has_spare) {
> > -               if (busiest->group_type > group_fully_busy) {
> > +               if ((busiest->group_type > group_fully_busy) &&
> > +                   (busiest->group_weight > 1)) {
> >                         /*
> >                          * If busiest is overloaded, try to fill 
> > spare
> >                          * capacity. This might end up creating spare 
> > capacity
> >
> >
> > When we calculate an imbalance at te smallest level, ie between CPUs 
> > (group_weight == 1),
> > we should try to spread tasks on cpus instead of trying to fill spare 
> > capacity.
>
> With this patch on top of v5.9, my latencies are unchanged.  I’m 
> building against current Linus now just in case I’m missing other 
> fixes.

I reran things to make sure the nothing changed on my test box this 
weekend:

5.4.0-rc1-00009-gfcf0553db6f4 (last good kernel)
Latency percentiles (usec) runtime 30 (s) (1000 total samples)
        50.0th: 180 (502 samples)
        75.0th: 227 (251 samples)
        90.0th: 268 (147 samples)
        95.0th: 300 (50 samples)
        *99.0th: 338 (41 samples)
        99.5th: 344 (4 samples)
        99.9th: 1186 (5 samples)
        min=25, max=1185

5.4.0-rc1-00010-g0b0695f2b34a (first bad kernel)
Latency percentiles (usec) runtime 150 (s) (960 total samples)
        50.0th: 166 (488 samples)
        75.0th: 210 (232 samples)
        90.0th: 254 (145 samples)
        95.0th: 299 (47 samples)
        *99.0th: 12688 (39 samples)
        99.5th: 13008 (5 samples)
        99.9th: 13104 (4 samples)
        min=24, max=13100

3650b228f83adda7e5ee532e2b90429c03f7b9ec (v5.10-rc1) + your patch

Latency percentiles (usec) runtime 30 (s) (1000 total samples)
        50.0th: 169 (505 samples)
        75.0th: 210 (246 samples)
        90.0th: 267 (151 samples)
        95.0th: 305 (48 samples)
        *99.0th: 12656 (40 samples)
        99.5th: 12944 (5 samples)
        99.9th: 13168 (5 samples)
        min=44, max=13155

-chris