Re: [PATCH] sched: wakeup buddy

From: Mike Galbraith
Date: Fri Mar 08 2013 - 01:44:28 EST


On Fri, 2013-03-08 at 10:37 +0800, Michael Wang wrote:
> On 03/07/2013 05:43 PM, Mike Galbraith wrote:
> > On Thu, 2013-03-07 at 09:36 +0100, Peter Zijlstra wrote:
> >> On Wed, 2013-03-06 at 15:06 +0800, Michael Wang wrote:
> >>
> >>> wake_affine() stuff is trying to bind related tasks closely, but it doesn't
> >>> work well according to the test on 'perf bench sched pipe' (thanks to Peter).
> >>
> >> so sched-pipe is a poor benchmark for this..
> >>
> >> Ideally we'd write a new benchmark that has some actual data footprint
> >> and we'd measure the cost of tasks being apart on the various cache
> >> metrics and see what affine wakeup does for it.
> >>
> >> Before doing something like what you're proposing, I'd have a hard look
> >> at WF_SYNC, it is possible we should disable/fix select_idle_sibling
> >> for sync wakeups.
> >
> > If nobody beats me to it, I'm going to try tracking shortest round trip
> > to idle, and use a multiple of that to shut select_idle_sibling() down.
> > If avg_idle approaches round trip time, there's no win to be had, we're
> > just wasting cycles.
>
> That's great if we have it, I'm a little doubt whether it is possible to
> find a better way to replace the select_idle_sibling() (look at the way
> it locates idle cpu...) in some cases, but I'm looking forward it ;-)

I'm not going to replace it, only stop it from wasting cycles when
there's very likely nothing to gain. Save task wakeup time, if delta
rq->clock - p->last_wakeup < N*shortest_idle or some such very cheap
metric. Wake ultra switchers L2 affine if allowed, only go hunting for
an idle L3 if the thing is on another package.

In general, I think things would work better if we'd just rate limit how
frequently we can wakeup migrate each individual task. We want
jabbering tasks to share L3, but we don't really want to trash L2 at an
awesome rate.

-Mike

--
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/