Re: [PATCH v4 2/2] sched/fair: Scan cluster before scanning LLC in wake-up path
From: Barry Song
Date: Fri Jun 10 2022 - 00:02:23 EST
On Fri, Jun 10, 2022 at 10:47 AM Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx> wrote:
>
> On Thu, 2022-06-09 at 20:06 +0800, Yicong Yang wrote:
> > From: Barry Song <song.bao.hua@xxxxxxxxxxxxx>
> >
> > For platforms having clusters like Kunpeng920, CPUs within the same cluster
> > have lower latency when synchronizing and accessing shared resources like
> > cache. Thus, this patch tries to find an idle cpu within the cluster of the
> > target CPU before scanning the whole LLC to gain lower latency.
> >
> > Note neither Kunpeng920 nor x86 Jacobsville supports SMT, so this patch
> > doesn't consider SMT for this moment.
> >
> > Testing has been done on Kunpeng920 by pinning tasks to one numa and two
> > numa. On Kunpeng920, Each numa has 8 clusters and each cluster has 4 CPUs.
> >
> > With this patch, We noticed enhancement on tbench within one numa or cross
> > two numa.
> >
> > On numa 0:
> > 5.19-rc1 patched
> > Hmean 1 350.27 ( 0.00%) 406.88 * 16.16%*
> > Hmean 2 702.01 ( 0.00%) 808.22 * 15.13%*
> > Hmean 4 1405.14 ( 0.00%) 1614.34 * 14.89%*
> > Hmean 8 2830.53 ( 0.00%) 3169.02 * 11.96%*
> > Hmean 16 5597.95 ( 0.00%) 6224.20 * 11.19%*
> > Hmean 32 10537.38 ( 0.00%) 10524.97 * -0.12%*
> > Hmean 64 8366.04 ( 0.00%) 8437.41 * 0.85%*
> > Hmean 128 7060.87 ( 0.00%) 7150.25 * 1.27%*
> >
> > On numa 0-1:
> > 5.19-rc1 patched
> > Hmean 1 346.11 ( 0.00%) 408.47 * 18.02%*
> > Hmean 2 693.34 ( 0.00%) 805.78 * 16.22%*
> > Hmean 4 1384.96 ( 0.00%) 1602.49 * 15.71%*
> > Hmean 8 2699.45 ( 0.00%) 3069.98 * 13.73%*
> > Hmean 16 5327.11 ( 0.00%) 5688.19 * 6.78%*
> > Hmean 32 10019.10 ( 0.00%) 11862.56 * 18.40%*
> > Hmean 64 13850.57 ( 0.00%) 17748.54 * 28.14%*
> > Hmean 128 12498.25 ( 0.00%) 15541.59 * 24.35%*
> > Hmean 256 11195.77 ( 0.00%) 13854.06 * 23.74%*
>
> Yicong,
>
> Have you tried any workload where tasks don't share data
> with each other but have sleep/wakeup? That's the case
> where we actually want to spread the tasks out among the clusters
> to void contention for L2 cache.
guess there isn't a way to universally win in term of the modifcation of
scheduling policies. The good news is that as long as LB is still there,
tasks still get a good chance to spread if they stop talking to each other
again and again after a small talk.
if tasks do like spreading but keep in constant touch with each other, we
may have to interfere with them from userspace by openMP, numactl or
taskset etc. as WAKE_AFFINE is always an enemy for this requirement.
>
> Will be nice to make sure there's no regression there for
> such workload.
>
> Code itself looks good.
>
> Reviewed-by: Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx>
>
> >
> > Tested-by: Yicong Yang <yangyicong@xxxxxxxxxxxxx>
> > Signed-off-by: Barry Song <song.bao.hua@xxxxxxxxxxxxx>
> > Signed-off-by: Yicong Yang <yangyicong@xxxxxxxxxxxxx>
> > ---
> > kernel/sched/fair.c | 44 +++++++++++++++++++++++++++++++++++++++++---
> > 1 file changed, 41 insertions(+), 3 deletions(-)
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 77b2048a9326..6d173e196ad3 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -6327,6 +6327,40 @@ static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd
> >
> > #endif /* CONFIG_SCHED_SMT */
> >
> > +#ifdef CONFIG_SCHED_CLUSTER
> > +/*
> > + * Scan the cluster domain for idle CPUs and clear cluster cpumask after scanning
> > + */
> > +static inline int scan_cluster(struct task_struct *p, struct cpumask *cpus,
> > + int target, int *nr)
> > +{
> > + struct sched_domain *sd = rcu_dereference(per_cpu(sd_cluster, target));
> > + int cpu, idle_cpu;
> > +
> > + /* TODO: Support SMT system with cluster topology */
> > + if (!sched_smt_active() && sd) {
> > + for_each_cpu_and(cpu, cpus, sched_domain_span(sd)) {
> > + if (!--*nr)
> > + break;
> > +
> > + idle_cpu = __select_idle_cpu(cpu, p);
> > + if ((unsigned int)idle_cpu < nr_cpumask_bits)
> > + return idle_cpu;
> > + }
> > +
> > + cpumask_andnot(cpus, cpus, sched_domain_span(sd));
> > + }
> > +
> > + return -1;
> > +}
> > +#else
> > +static inline int scan_cluster(struct task_struct *p, struct cpumask *cpus,
> > + int target, int *nr)
> > +{
> > + return -1;
> > +}
> > +#endif
> > +
> > /*
> > * Scan the LLC domain for idle CPUs; this is dynamically regulated by
> > * comparing the average scan cost (tracked in sd->avg_scan_cost) against the
> > @@ -6375,6 +6409,10 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
> > time = cpu_clock(this);
> > }
> >
> > + idle_cpu = scan_cluster(p, cpus, target, &nr);
> > + if ((unsigned int)idle_cpu < nr_cpumask_bits)
> > + return idle_cpu;
> > +
> > for_each_cpu_wrap(cpu, cpus, target + 1) {
> > if (has_idle_core) {
> > i = select_idle_core(p, cpu, cpus, &idle_cpu);
> > @@ -6382,7 +6420,7 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
> > return i;
> >
> > } else {
> > - if (!--nr)
> > + if (--nr <= 0)
> > return -1;
> > idle_cpu = __select_idle_cpu(cpu, p);
> > if ((unsigned int)idle_cpu < nr_cpumask_bits)
> > @@ -6481,7 +6519,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
> > /*
> > * If the previous CPU is cache affine and idle, don't be stupid:
> > */
> > - if (prev != target && cpus_share_cache(prev, target) &&
> > + if (prev != target && cpus_share_resources(prev, target) &&
> > (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
> > asym_fits_capacity(task_util, prev))
> > return prev;
> > @@ -6507,7 +6545,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
> > p->recent_used_cpu = prev;
> > if (recent_used_cpu != prev &&
> > recent_used_cpu != target &&
> > - cpus_share_cache(recent_used_cpu, target) &&
> > + cpus_share_resources(recent_used_cpu, target) &&
> > (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)) &&
> > cpumask_test_cpu(p->recent_used_cpu, p->cpus_ptr) &&
> > asym_fits_capacity(task_util, recent_used_cpu)) {
>
Thanks
Barry