On Mon, Feb 07, 2022 at 11:40:13AM +0800, Chen Yu wrote:
> It would be ideal to have a crystal ball to predict the success rate
> of finding an idle cpu/core in the LLC. If it is doomed to fail,
> there is no need to search in the LLC domain. There are many potential
> metrics which could be used to predict the success rate. And the
> metric should be carefully chosen that, it should help reduce the
> unnecessary cpu runqueue scan, but meanwhile not give up the opportunity
> to find an idle cpu.
>
> Choose average cpu utilization as the candidate, since the util_avg is
> a metric of accumulated historic activity, which seems to be more accurate
> than instantaneous metrics(such as rq->nr_running) on calculating the probability
> of find an idle cpu. Only when the average cpu utilization has reaches
> 85% of the total cpu capacity, this domain is regarded as overloaded.
> The reason to choose 85% is that, this is the threshold of an overloaded
> LLC sched group(imbalance_pct = 117, threshold = 100 / 117 * 100%).
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 5146163bfabb..1a58befe892d 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -6280,6 +6281,10 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
> if (!this_sd)
> return -1;
>
> + sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
> + if (sd_share && READ_ONCE(sd_share->overloaded))
> + return -1;
> +
> cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
>
> if (sched_feat(SIS_PROP) && !has_idle_core) {
> @@ -9268,6 +9275,29 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd
> WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
> trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
> }
> +
> + /*
> + * Check if the LLC domain is overloaded. The overload hint
> + * could be used to skip the LLC domain idle cpu search in
> + * select_idle_cpu(). The update of this hint occurs during
> + * periodic load balancing, rather than frequent newidle balance.
> + */
> + if (env->idle != CPU_NEWLY_IDLE &&
> + env->sd->span_weight == per_cpu(sd_llc_size, env->dst_cpu)) {
> + struct sched_domain_shared *sd_share =
> + rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
> +
> + if (!sd_share)
> + return;
> +
> + /*
> + * Derived from group_is_overloaded(). The default imbalance_pct
> + * is 117 on LLC domain, which means the threshold of average
> + * utilization is 85%.
> + */
> + WRITE_ONCE(sd_share->overloaded, (sds->total_capacity * 100) <
> + (sum_util * env->sd->imbalance_pct));
> + }
> }
So the only problem I have with this is that this is somewhat of a
binary toggle. The moment we hit that magical 85% we suddenly change
behaviour.
Would it not be possible to replace the SIS_PROP logic with something
based on this sum_util metric? Such that when sum_util is very low we
scan more, while when sum_util hits 85% we naturally stop scanning
entirely.
That way the behaviour is gradual.
On Mon, Feb 07, 2022 at 02:52:53PM +0100, Peter Zijlstra wrote:
> On Mon, Feb 07, 2022 at 11:40:13AM +0800, Chen Yu wrote:
> > It would be ideal to have a crystal ball to predict the success rate
> > of finding an idle cpu/core in the LLC. If it is doomed to fail,
> > there is no need to search in the LLC domain. There are many potential
> > metrics which could be used to predict the success rate. And the
> > metric should be carefully chosen that, it should help reduce the
> > unnecessary cpu runqueue scan, but meanwhile not give up the opportunity
> > to find an idle cpu.
> >
> > Choose average cpu utilization as the candidate, since the util_avg is
> > a metric of accumulated historic activity, which seems to be more accurate
> > than instantaneous metrics(such as rq->nr_running) on calculating the probability
> > of find an idle cpu. Only when the average cpu utilization has reaches
> > 85% of the total cpu capacity, this domain is regarded as overloaded.
> > The reason to choose 85% is that, this is the threshold of an overloaded
> > LLC sched group(imbalance_pct = 117, threshold = 100 / 117 * 100%).
>
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 5146163bfabb..1a58befe892d 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
>
> > @@ -6280,6 +6281,10 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
> > if (!this_sd)
> > return -1;
> >
> > + sd_share = rcu_dereference(per_cpu(sd_llc_shared, target));
> > + if (sd_share && READ_ONCE(sd_share->overloaded))
> > + return -1;
> > +
> > cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
> >
> > if (sched_feat(SIS_PROP) && !has_idle_core) {
>
> > @@ -9268,6 +9275,29 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd
> > WRITE_ONCE(rd->overutilized, SG_OVERUTILIZED);
> > trace_sched_overutilized_tp(rd, SG_OVERUTILIZED);
> > }
> > +
> > + /*
> > + * Check if the LLC domain is overloaded. The overload hint
> > + * could be used to skip the LLC domain idle cpu search in
> > + * select_idle_cpu(). The update of this hint occurs during
> > + * periodic load balancing, rather than frequent newidle balance.
> > + */
> > + if (env->idle != CPU_NEWLY_IDLE &&
> > + env->sd->span_weight == per_cpu(sd_llc_size, env->dst_cpu)) {
> > + struct sched_domain_shared *sd_share =
> > + rcu_dereference(per_cpu(sd_llc_shared, env->dst_cpu));
> > +
> > + if (!sd_share)
> > + return;
> > +
> > + /*
> > + * Derived from group_is_overloaded(). The default imbalance_pct
> > + * is 117 on LLC domain, which means the threshold of average
> > + * utilization is 85%.
> > + */
> > + WRITE_ONCE(sd_share->overloaded, (sds->total_capacity * 100) <
> > + (sum_util * env->sd->imbalance_pct));
> > + }
> > }
>
> So the only problem I have with this is that this is somewhat of a
> binary toggle. The moment we hit that magical 85% we suddenly change
> behaviour.
>
> Would it not be possible to replace the SIS_PROP logic with something
> based on this sum_util metric? Such that when sum_util is very low we
> scan more, while when sum_util hits 85% we naturally stop scanning
> entirely.
>
> That way the behaviour is gradual.
Ok, let me do some evaluation on this, maybe something like nr_to_scan = f(sum_util)
thanks,
Chenyu