Re: [patch 5/6] sched: disable sched feature TTWU_QUEUE by default

[Mike Galbraith]

On Fri, 2011-11-18 at 15:03 -0800, Suresh Siddha wrote:
> plain text document attachment (disable_sched_ttwu_queue.patch)
> Context-switch intensive microbenchmark on a 8-socket system had
> ~600K times more resched IPI's on each logical CPU with this feature enabled
> by default. Disabling this features makes that microbenchmark perform 5 times
> better.
> 
> Also disabling this feature showed 2% performance improvement on a 8-socket
> OLTP workload.
> 
> More heurestics are needed when and how to use this feature by default.
> For now, disable it by default.

Yeah, the overhead for very hefty switchers is high enough to increase
TCP_RR latency up to 13% in my testing.  I used a trylock() to generally
not eat that, but leave the contended case improvement intact.

Peter suggested trying doing the IPI only when crossing cache
boundaries, which worked for me as well.

---
 kernel/sched.c |   24 +++++++++++++++++++++++-
 1 file changed, 23 insertions(+), 1 deletion(-)

Index: linux-3.0-tip/kernel/sched.c
===================================================================
--- linux-3.0-tip.orig/kernel/sched.c
+++ linux-3.0-tip/kernel/sched.c
@@ -2784,12 +2784,34 @@ static int ttwu_activate_remote(struct t
 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
 #endif /* CONFIG_SMP */
 
+static int ttwu_share_cache(int this_cpu, int cpu)
+{
+#ifndef CONFIG_X86
+	struct sched_domain *sd;
+	int ret = 0;
+
+	rcu_read_lock();
+	for_each_domain(this_cpu, sd) {
+		if (!cpumask_test_cpu(cpu, sched_domain_span(sd)))
+			continue;
+
+		ret = (sd->flags & SD_SHARE_PKG_RESOURCES);
+		break;
+	}
+	rcu_read_unlock();
+
+	return ret;
+#else
+	return per_cpu(cpu_llc_id, this_cpu) == per_cpu(cpu_llc_id, cpu);
+#endif
+}
+
 static void ttwu_queue(struct task_struct *p, int cpu)
 {
 	struct rq *rq = cpu_rq(cpu);
 
 #if defined(CONFIG_SMP)
-	if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
+	if (sched_feat(TTWU_QUEUE) && !ttwu_share_cache(smp_processor_id(), cpu)) {
 		sched_clock_cpu(cpu); /* sync clocks x-cpu */
 		ttwu_queue_remote(p, cpu);
 		return;
  

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/