Re: [patch v3] sched: fix select_idle_sibling() induced bouncing

[Mike Galbraith <efault@gmx.de>]

On Mon, 2012-06-11 at 11:53 -0700, Suresh Siddha wrote: 
> On Mon, 2012-06-11 at 19:55 +0200, Mike Galbraith wrote:
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -5928,6 +5928,11 @@ static void destroy_sched_domains(struct
> >   * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
> >   * allows us to avoid some pointer chasing select_idle_sibling().
> >   *
> > + * Iterate domains and sched_groups upward, assigning CPUs to be
> 
> You are actually iterating downwards (starting from the highest domain
> with the SHARE_PKG_RESOURCES flag) in the patch.

(not _my_ fault everybody else on the planet sees them upside-down;)

> > + * select_idle_sibling() hw buddy.  Cross-wiring hw makes bouncing
> > + * due to random perturbation self canceling, ie sw buddies pull
> > + * their counterpart to their CPU's hw counterpart.
> > + *
> 
> Also it will be nice to include all the data you have observed as part
> of the changelog.

Comment inverted, worst pain encounter numbers added.

sched: fix select_idle_sibling() induced bouncing

Traversing an entire package is not only expensive, it also leads to tasks
bouncing all over a partially idle and possible quite large package.  Fix
that up by assigning a 'buddy' CPU to try to motivate.  Each buddy may try
to motivate that one other CPU, if it's busy, tough, it may then try it's
SMT sibling, but that's all this optimization is allowed to cost.

Sibling cache buddies are cross-wired to prevent bouncing.

4 socket 40 core + SMT Westmere box, single 30 sec tbench runs
clients     1       2       4        8       16       32       64      128
pre        30      41     118      645     3769     6214    12233    14312
post      299     603    1211     2418     4697     6847    11606    14557

Signed-off-by: Mike Galbraith <efault@gmx.de>

---
 include/linux/sched.h |    1 +
 kernel/sched/core.c   |   39 ++++++++++++++++++++++++++++++++++++++-
 kernel/sched/fair.c   |   28 +++++++---------------------
 3 files changed, 46 insertions(+), 22 deletions(-)

--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -955,6 +955,7 @@ struct sched_domain {
 	unsigned int smt_gain;
 	int flags;			/* See SD_* */
 	int level;
+	int idle_buddy;			/* cpu assigned to select_idle_sibling() */
 
 	/* Runtime fields. */
 	unsigned long last_balance;	/* init to jiffies. units in jiffies */
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -5928,6 +5928,11 @@ static void destroy_sched_domains(struct
  * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
  * allows us to avoid some pointer chasing select_idle_sibling().
  *
+ * Iterate domains and sched_groups downward, assigning CPUs to be
+ * select_idle_sibling() hw buddy.  Cross-wiring hw makes bouncing
+ * due to random perturbation self canceling, ie sw buddies pull
+ * their counterpart to their CPU's hw counterpart.
+ *
  * Also keep a unique ID per domain (we use the first cpu number in
  * the cpumask of the domain), this allows us to quickly tell if
  * two cpus are in the same cache domain, see cpus_share_cache().
@@ -5943,8 +5948,40 @@ static void update_domain_cache(int cpu)
 	int id = cpu;
 
 	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
-	if (sd)
+	if (sd) {
+		struct sched_domain *tmp = sd;
+		struct sched_group *sg, *prev;
+		bool right;
+
+		/*
+		 * Traversse to first CPU in group, and count hops
+		 * to cpu from there, switching direction on each
+		 * hop, never ever pointing the last CPU rightward.
+		 */
+		do {
+			id = cpumask_first(sched_domain_span(tmp));
+			prev = sg = tmp->groups;
+			right = 1;
+
+			while (cpumask_first(sched_group_cpus(sg)) != id)
+				sg = sg->next;
+
+			while (!cpumask_test_cpu(cpu, sched_group_cpus(sg))) {
+				prev = sg;
+				sg = sg->next;
+				right = !right;
+			}
+
+			/* A CPU went down, never point back to domain start. */
+			if (right && cpumask_first(sched_group_cpus(sg->next)) == id)
+				right = false;
+
+			sg = right? sg->next : prev;
+			tmp->idle_buddy = cpumask_first(sched_group_cpus(sg));
+		} while ((tmp = tmp->child));
+
 		id = cpumask_first(sched_domain_span(sd));
+	}
 
 	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
 	per_cpu(sd_llc_id, cpu) = id;
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -2642,8 +2642,6 @@ static int select_idle_sibling(struct ta
 	int cpu = smp_processor_id();
 	int prev_cpu = task_cpu(p);
 	struct sched_domain *sd;
-	struct sched_group *sg;
-	int i;
 
 	/*
 	 * If the task is going to be woken-up on this cpu and if it is
@@ -2660,29 +2658,17 @@ static int select_idle_sibling(struct ta
 		return prev_cpu;
 
 	/*
-	 * Otherwise, iterate the domains and find an elegible idle cpu.
+	 * Otherwise, check assigned siblings to find an elegible idle cpu.
 	 */
 	sd = rcu_dereference(per_cpu(sd_llc, target));
+
 	for_each_lower_domain(sd) {
-		sg = sd->groups;
-		do {
-			if (!cpumask_intersects(sched_group_cpus(sg),
-						tsk_cpus_allowed(p)))
-				goto next;
-
-			for_each_cpu(i, sched_group_cpus(sg)) {
-				if (!idle_cpu(i))
-					goto next;
-			}
-
-			target = cpumask_first_and(sched_group_cpus(sg),
-					tsk_cpus_allowed(p));
-			goto done;
-next:
-			sg = sg->next;
-		} while (sg != sd->groups);
+		if (!cpumask_test_cpu(sd->idle_buddy, tsk_cpus_allowed(p)))
+			continue;
+		if (idle_cpu(sd->idle_buddy))
+			return sd->idle_buddy;
 	}
-done:
+
 	return target;
 }