Re: [performance bug] volanomark regression on 37-rc1

[Mike Galbraith <efault@gmx.de>]

On Tue, 2010-11-16 at 08:31 -0800, Nikhil Rao wrote:
> On Tue, Nov 16, 2010 at 7:26 AM, Mike Galbraith <efault@gmx.de> wrote:
> > On Tue, 2010-11-16 at 20:38 +0600, Rakib Mullick wrote:
> >
> >> Does VolanoMark is used for scheduler benchmarking? If I'm not wrong,
> >> I don't think it directly relates to scheduler benchmarking.
> >
> > It's not generally considered to be a wonderful benchmark, but it is a
> > good indicator, and worth keeping an eye on IMHO.
> >
> > I don't recall whether that patch works with the idle testcase without
> > resetting the throttle, or if it's only a bit less effective.  If it's
> > only a little less effective, I'd be inclined to just whack the reset as
> > Alex did.  Whatever is done has to prevent high frequency balancing.
> >
> 
> >From what I recall, I think removing the reset makes the original
> patch a little less effective. I agree that we can remove the reset if
> it hurts high frequency balancing.

Ok, let's do that.  I added your ack, OK?

From: Alex Shi <alex.shi@intel.com>
Date: Tue, 16 Nov 2010 17:34:02 +0800

    sched: volanomark regression fix

    Commit fab4762 triggers excessive idle balancing, causing a ~30% loss in
    volanomark throughput.  Remove idle balancing throttle reset.

    Signed-off-by: Mike Galbraith <efault@gmx.de>
    Acked-by: Nikhil Rao <ncrao@google.com>
    Reported-by: Alex Shi <alex.shi@intel.com>
    Cc: Ingo Molnar <mingo@elte.hu>
    Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
    LKML-Reference: <new-submission>

---
 kernel/sched_fair.c |    4 ----
 1 file changed, 4 deletions(-)

Index: linux-2.6/kernel/sched_fair.c
===================================================================
--- linux-2.6.orig/kernel/sched_fair.c
+++ linux-2.6/kernel/sched_fair.c
@@ -1758,10 +1758,6 @@ static void pull_task(struct rq *src_rq,
 	set_task_cpu(p, this_cpu);
 	activate_task(this_rq, p, 0);
 	check_preempt_curr(this_rq, p, 0);
-
-	/* re-arm NEWIDLE balancing when moving tasks */
-	src_rq->avg_idle = this_rq->avg_idle = 2*sysctl_sched_migration_cost;
-	this_rq->idle_stamp = 0;
 }
 
 /*