[PATCH 0/4] Reduce scanning of runqueues in select_idle_sibling

[Mel Gorman]

On Fri, Dec 11, 2020 at 10:51:17AM +0100, Vincent Guittot wrote:
> On Thu, 10 Dec 2020 at 12:04, Mel Gorman <mgorman at techsingularity.net> wrote:
> >
> > On Thu, Dec 10, 2020 at 10:38:37AM +0100, Vincent Guittot wrote:
> > > > while testing your patchset and Aubrey one on top of tip, I'm facing
> > > > some perf regression on my arm64 numa system on hackbench and reaim.
> > > > The regression seems to comes from your patchset but i don't know
> > > > which patch in particular yet
> > > >
> > > > hackbench -l 256000 -g 1
> > > >
> > > > v5.10-rc7 + tip/sched/core 13,255(+/- 3.22%)
> > > > with your patchset         15.368(+/- 2.74)  -15.9%
> > > >
> > > > I'm also seeing perf regression on reaim but this one needs more
> > > > investigation before confirming
> > > >
> > > > TBH, I was not expecting regressions. I'm running more test to find
> > > > which patch is the culprit
> > >
> > > The regression comes from patch 3: sched/fair: Do not replace
> > > recent_used_cpu with the new target
> > >
> >
> > That's not entirely surprising. The intent of the patch is to increase the
> > hit rate of p->recent_used_cpu but it's not a guaranteed win due to two
> > corner cases. If multiple tasks have the same p->recent_used_cpu, they can
> > race to use that CPU and stack as a result instead of searching the domain.
> > If SMT is enabled then p->recent_used_cpu can point to an idle CPU that has
> > a busy sibling which the search would have avoided in select_idle_core().
> >
> > I think you are using processes and sockets for hackbench but as you'll
> > see later, hackbench can be used both to show losses and gains.
> 
> I run more hackbench tests with pipe and socket and both show
> regression with patch 3 whereas this is significant improvement with
> other patches and Aubrey's one
> 

Is SMT enabled on your test machine? If not, then patch 4 should make no
difference but if SMT is enabled, I wonder how this untested version of
patch 3 behaves for you. The main difference is that the recent used cpu
is used as a search target so that it would still check if it's an idle
core and if not, fall through so it's used as an idle CPU after checking
it's allowed by p->cpus_ptr.

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 5c41875aec23..63980bcf6e70 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6275,21 +6275,14 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
 		return prev;
 	}
 
-	/* Check a recently used CPU as a potential idle candidate: */
+	/* Check a recently used CPU as a search target: */
 	recent_used_cpu = p->recent_used_cpu;
+	p->recent_used_cpu = prev;
 	if (recent_used_cpu != prev &&
 	    recent_used_cpu != target &&
 	    cpus_share_cache(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)) {
-		/*
-		 * Replace recent_used_cpu with prev as it is a potential
-		 * candidate for the next wake:
-		 */
-		p->recent_used_cpu = prev;
-		return recent_used_cpu;
-	}
+	    (available_idle_cpu(recent_used_cpu) || sched_idle_cpu(recent_used_cpu)))
+		target = recent_used_cpu;
 
 	/*
 	 * For asymmetric CPU capacity systems, our domain of interest is
@@ -6768,9 +6761,6 @@ select_task_rq_fair(struct task_struct *p, int prev_cpu, int wake_flags)
 	} else if (wake_flags & WF_TTWU) { /* XXX always ? */
 		/* Fast path */
 		new_cpu = select_idle_sibling(p, prev_cpu, new_cpu);
-
-		if (want_affine)
-			current->recent_used_cpu = cpu;
 	}
 	rcu_read_unlock();
 

-- 
Mel Gorman
SUSE Labs