Re: [PATCH 2/3] sched: Move idle_balance() to post_schedule
From: Steven Rostedt
Date: Thu Feb 14 2013 - 09:25:35 EST
On Wed, 2013-02-13 at 19:43 +0100, Peter Zijlstra wrote:
> How does that follow? We can have to-idle switches _far_ more often than
> we balance.
I ran "perf stat -a -r 100 hackbench 500" on an i7 4 core hyperthreaded
box, and got the following results:
With no patches applied:
Performance counter stats for '/work/c/hackbench 500' (100 runs):
199820.045583 task-clock # 8.016 CPUs utilized ( +- 5.29% ) [100.00%]
3,594,264 context-switches # 0.018 M/sec ( +- 5.94% ) [100.00%]
352,240 cpu-migrations # 0.002 M/sec ( +- 3.31% ) [100.00%]
1,006,732 page-faults # 0.005 M/sec ( +- 0.56% )
293,801,912,874 cycles # 1.470 GHz ( +- 4.20% ) [100.00%]
261,808,125,109 stalled-cycles-frontend # 89.11% frontend cycles idle ( +- 4.38% ) [100.00%]
<not supported> stalled-cycles-backend
135,521,344,089 instructions # 0.46 insns per cycle
# 1.93 stalled cycles per insn ( +- 4.37% ) [100.00%]
26,198,116,586 branches # 131.109 M/sec ( +- 4.59% ) [100.00%]
115,326,812 branch-misses # 0.44% of all branches ( +- 4.12% )
24.929136087 seconds time elapsed ( +- 5.31% )
With the two idle_balance patches applied:
Performance counter stats for '/work/c/hackbench 500' (100 runs):
178619.929457 task-clock # 8.011 CPUs utilized ( +- 4.16% ) [100.00%]
3,171,229 context-switches # 0.018 M/sec ( +- 3.30% ) [100.00%]
323,115 cpu-migrations # 0.002 M/sec ( +- 2.47% ) [100.00%]
994,506 page-faults # 0.006 M/sec ( +- 0.52% )
269,744,391,573 cycles # 1.510 GHz ( +- 2.12% ) [100.00%]
238,589,242,461 stalled-cycles-frontend # 88.45% frontend cycles idle ( +- 2.26% ) [100.00%]
<not supported> stalled-cycles-backend
124,298,251,712 instructions # 0.46 insns per cycle
# 1.92 stalled cycles per insn ( +- 1.99% ) [100.00%]
23,918,305,712 branches # 133.906 M/sec ( +- 2.13% ) [100.00%]
105,863,415 branch-misses # 0.44% of all branches ( +- 2.14% )
22.296151996 seconds time elapsed ( +- 4.18% )
And finally with the patch at the bottom of this email applied:
Performance counter stats for '/work/c/hackbench 500' (100 runs):
170170.547682 task-clock # 8.021 CPUs utilized ( +- 5.59% ) [100.00%]
3,118,923 context-switches # 0.018 M/sec ( +- 4.82% ) [100.00%]
318,479 cpu-migrations # 0.002 M/sec ( +- 2.58% ) [100.00%]
988,187 page-faults # 0.006 M/sec ( +- 0.62% )
271,343,352,987 cycles # 1.595 GHz ( +- 3.84% ) [100.00%]
240,599,089,430 stalled-cycles-frontend # 88.67% frontend cycles idle ( +- 4.24% ) [100.00%]
<not supported> stalled-cycles-backend
125,888,645,748 instructions # 0.46 insns per cycle
# 1.91 stalled cycles per insn ( +- 3.97% ) [100.00%]
24,219,147,811 branches # 142.323 M/sec ( +- 4.22% ) [100.00%]
105,077,636 branch-misses # 0.43% of all branches ( +- 3.70% )
21.214840224 seconds time elapsed ( +- 5.61% )
Yeah, it seems the extra check for rq empty helps. But even without
that, the current idle patches still seem pretty good.
-- Steve
diff --git a/kernel/sched/idle_task.c b/kernel/sched/idle_task.c
index 66b5220..025350b 100644
--- a/kernel/sched/idle_task.c
+++ b/kernel/sched/idle_task.c
@@ -16,7 +16,9 @@ select_task_rq_idle(struct task_struct *p, int sd_flag, int flags)
static void post_schedule_idle(struct rq *rq)
{
- idle_balance(smp_processor_id(), rq);
+ /* rq lock was released, make sure system is still idle */
+ if (likely(!rq->nr_running))
+ idle_balance(smp_processor_id(), rq);
}
#endif /* CONFIG_SMP */
/*
--
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/