Re: [PATCH 2/2] sched/fair: Scan cluster before scanning LLC in wake-up path
From: Barry Song
Date: Wed Dec 22 2021 - 17:36:19 EST
On Thu, Dec 23, 2021 at 2:27 AM Vincent Guittot
<vincent.guittot@xxxxxxxxxx> wrote:
>
> On Wed, 15 Dec 2021 at 05:11, Yicong Yang <yangyicong@xxxxxxxxxxxxx> 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. 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:
> > 5.16-rc1 patched
> > Hmean 1 329.17 ( 0.00%) 384.84 * 16.91%*
> > Hmean 2 654.09 ( 0.00%) 768.77 * 17.53%*
> > Hmean 4 1321.41 ( 0.00%) 1538.10 * 16.40%*
> > Hmean 8 2650.43 ( 0.00%) 3048.86 * 15.03%*
> > Hmean 16 5322.37 ( 0.00%) 5975.20 * 12.27%*
> > Hmean 32 10002.11 ( 0.00%) 10085.57 * 0.83%*
> > Hmean 64 7910.39 ( 0.00%) 7936.37 * 0.33%*
> > Hmean 128 6745.25 ( 0.00%) 6811.28 * 0.98%*
> >
> > On numa 0-1:
> > 5.16-rc1 patched
> > Hmean 1 326.61 ( 0.00%) 385.36 * 17.99%*
> > Hmean 2 650.62 ( 0.00%) 770.57 * 18.44%*
> > Hmean 4 1318.05 ( 0.00%) 1534.83 * 16.45%*
> > Hmean 8 2621.50 ( 0.00%) 3030.10 * 15.59%*
> > Hmean 16 5252.17 ( 0.00%) 6023.08 * 14.68%*
> > Hmean 32 9829.30 ( 0.00%) 9856.33 * 0.28%*
> > Hmean 64 12452.66 ( 0.00%) 17338.48 * 39.24%*
> > Hmean 128 14181.24 ( 0.00%) 15025.24 * 5.95%*
> > Hmean 256 12239.07 ( 0.00%) 13080.16 * 6.87%*
> > Hmean 512 14297.00 ( 0.00%) 15063.76 * 5.36%*
> >
> > This will also help to improve the MySQL. With MySQL server
> > running on numa 0 and client running on numa 1, both QPS and
> > latency is imporved on read-write case:
> > 5.16-rc1 patched
> > QPS-24threads 195327.48 202081.28(+3.46%)
> > QPS-32threads 242039.4 247059.5(+2.07%)
> > QPS-64threads 243024.52 254274.47(+4.63%)
> > avg-lat-24threads 2.46 2.37(+3.66%)
> > avg-lat-36threads 2.64 2.59(+1.89%)
> > avg-lat-64threads 5.27 5.03(+4.55%)
> >
> > 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 | 46 +++++++++++++++++++++++++++++++++++++++++----
> > 1 file changed, 42 insertions(+), 4 deletions(-)
> >
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 6e476f6d9435..8a5795c78af8 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -6230,12 +6230,46 @@ 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, int prev_cpu, int target)
> > +{
> > + struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
> > + struct sched_domain *sd = rcu_dereference(per_cpu(sd_cluster, target));
> > + int cpu, idle_cpu;
> > +
> > + /* TODO: Support SMT case while a machine with both cluster and SMT born */
> > + if (!sched_smt_active() && sd) {
> > + for_each_cpu_and(cpu, cpus, sched_domain_span(sd)) {
> > + idle_cpu = __select_idle_cpu(cpu, p);
> > + if ((unsigned int)idle_cpu < nr_cpumask_bits)
> > + return idle_cpu;
> > + }
> > +
> > + /* Don't ping-pong tasks in and out cluster frequently */
> > + if (cpus_share_cluster(target, prev_cpu))
> > + return target;
> > +
> > + cpumask_andnot(cpus, cpus, sched_domain_span(sd));
> > + }
> > +
> > + return -1;
> > +}
> > +#else
> > +static inline int scan_cluster(struct task_struct *p, int prev_cpu, int target)
> > +{
> > + 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
> > * average idle time for this rq (as found in rq->avg_idle).
> > */
> > -static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int target)
> > +static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool has_idle_core, int prev_cpu, int target)
> > {
> > struct cpumask *cpus = this_cpu_cpumask_var_ptr(select_idle_mask);
> > int i, cpu, idle_cpu = -1, nr = INT_MAX;
> > @@ -6250,6 +6284,10 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
> >
> > cpumask_and(cpus, sched_domain_span(sd), p->cpus_ptr);
> >
> > + idle_cpu = scan_cluster(p, prev_cpu, target);
>
> Why don't you do the above before calling select_idle_cpu() like for smt ?
yep. I did make two versions :
1. scan_cluster inside select_idle_cpu()
2. scan_cluster before select_idle_cpu()
Then I felt 1 looks a bit better than 2 as 2 requires changes in two
places. We still need to change
select_idle_cpu() to do cpumask_andnot() to exclude the cluster as we
have scanned the cluster
before select_idle_cpu(). In the first way, only one place is modified.
>
> > + if ((unsigned int)idle_cpu < nr_cpumask_bits)
> > + return idle_cpu;
> > +
> > if (sched_feat(SIS_PROP) && !has_idle_core) {
> > u64 avg_cost, avg_idle, span_avg;
> > unsigned long now = jiffies;
> > @@ -6384,7 +6422,7 @@ 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_cluster(prev, target) &&
>
> This is misleading because cpus_share_cluster is meaningless for most of system
yep. The name is misleading though it hasn't side effects for
non-cluster systems as
cpus_share_cluster equals cpus_share_cache for them.
It seems this can be much better if we do the below as your comments
in patch1/2:
+bool cpus_share_resources(int this_cpu, int that_cpu)
+{
+ if (this_cpu == that_cpu)
+ return true;
+
+ return per_cpu(sd_share_id, this_cpu) == per_cpu(sd_share_id, that_cpu);
+}
+
which returns true when cpu shares resources which can be LLC or cluster
>
> Then, you don't care at all of the llc if there is a cluster domain ?
Basically, what I have seen is that in tbench, target is likely to be
idle especially when
the system load is not full. and waker and wakee are unlikely to run
in the same cluster
as fork has splitted them for LB. so we will almost always lose the
chance to pull them
together forever. This change made a big gain for tbench especially
for light and medium
load.
But Yicong did see some negative impact on benchmarks like pgbench. so
this change
is not a universal win. The difficulty might be how to make a decision
which can be
more aware of the hierarchy of two levels - cluster and llc, and aware
of the requirements
of the workload.
>
>
> > (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
> > asym_fits_capacity(task_util, prev))
> > return prev;
> > @@ -6408,7 +6446,7 @@ 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_cluster(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)) {
> > @@ -6449,7 +6487,7 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
> > }
> > }
> >
> > - i = select_idle_cpu(p, sd, has_idle_core, target);
> > + i = select_idle_cpu(p, sd, has_idle_core, prev, target);
> > if ((unsigned)i < nr_cpumask_bits)
> > return i;
> >
> > --
> > 2.33.0
> >
Thanks
Barry