Re: [patch 3/6] sched, nohz: sched group, domain aware nohz idle load balancing

[Peter Zijlstra <peterz@infradead.org>]

On Fri, 2011-11-18 at 15:03 -0800, Suresh Siddha wrote:
>  static inline int nohz_kick_needed(struct rq *rq, int cpu)
>  {
>         unsigned long now = jiffies;
>         struct sched_domain *sd;
>  
> +       if (unlikely(idle_cpu(cpu)))
> +               return 0;
> +
>         /*
>          * We were recently in tickless idle mode. At the first busy tick
>          * after returning from idle, we will update the busy stats.
> @@ -5120,36 +5047,43 @@ static inline int nohz_kick_needed(struc
>         if (unlikely(test_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu)))) {
>                 clear_bit(NOHZ_TICK_STOPPED, nohz_flags(cpu));
>  
> +               cpumask_clear_cpu(cpu, nohz.idle_cpus_mask);
> +               atomic_dec(&nohz.nr_cpus);
> +
>                 for_each_domain(cpu, sd)
>                         atomic_inc(&sd->groups->sgp->nr_busy_cpus);
>         }
>  
> +       /*
> +        * None are in tickless mode and hence no need for NOHZ idle load
> +        * balancing.
> +        */
> +       if (likely(!atomic_read(&nohz.nr_cpus)))
>                 return 0;
>  
> +       if (time_before(now, nohz.next_balance))
>                 return 0;
>  
> +       if (rq->nr_running >= 2)
> +               goto need_kick;
>  
> +       for_each_domain(cpu, sd) {
> +               struct sched_group *sg = sd->groups;
> +               struct sched_group_power *sgp = sg->sgp;
> +               int nr_busy = atomic_read(&sgp->nr_busy_cpus);
> +
> +               if (nr_busy > 1 && (nr_busy * SCHED_LOAD_SCALE > sgp->power))
> +                       goto need_kick;

This looks wrong, its basically always true for a box with HT.

sgp->power is a measure of how much compute power this group has, its
basic form is sg->weight * SCHED_POWER_SCALE and is reduced from there;
HT siblings get less since they're not as powerful as two actual cores
and we deduct time spend on RT-tasks and IRQs etc..

So how does comparing the load of non-nohz cpus to that make sense?

> +
> +               if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
> +                   && (cpumask_first_and(nohz.idle_cpus_mask,
> +                                         sched_domain_span(sd)) < cpu))
> +                       goto need_kick;
>         }
> +
>         return 0;
> +need_kick:
> +       return 1;
>  }