Messages in this thread | | | From | bsegall@google ... | Subject | Re: [PATCH 2/2] sched: Rewrite per entity runnable load average tracking | Date | Wed, 09 Jul 2014 12:07:08 -0700 |
| |
Peter Zijlstra <peterz@infradead.org> writes:
> On Wed, Jul 09, 2014 at 09:07:53AM +0800, Yuyang Du wrote: >> That is chalenging... Can someone (Peter) grant us a lock of the remote rq? :) > > Nope :-).. we got rid of that lock for a good reason. > > Also, this is one area where I feel performance really trumps > correctness, we can fudge the blocked load a little. So the > sched_clock_cpu() difference is a strict upper bound on the > rq_clock_task() difference (and under 'normal' circumstances shouldn't > be much off).
Well, unless IRQ_TIME_ACCOUNTING or such is on, in which case you lose. Or am I misunderstanding the suggestion? Actually the simplest thing would probably be to grab last_update_time (which on 32-bit could be done with the _copy hack) and use that. Then I think the accuracy is only worse than current in that you can lose runnable load as well as blocked load, and that it isn't as easily corrected - currently if the blocked tasks wake up they'll add the correct numbers to runnable_load_avg, even if blocked_load_avg is screwed up and hit zero. This code would have to wait until it stabilized again.
> > So we could simply use a timestamps from dequeue and one from enqueue, > and use that. > > As to the remote subtraction, a RMW on another cacheline than the > rq->lock one should be good; esp since we don't actually observe the > per-rq total often (once per tick or so) I think, no?
Yeah, it's definitely a different cacheline, and the current code only reads per-ms or on loadbalance migration.
> > The thing is, we do not want to disturb scheduling on whatever cpu the > task last ran on if we wake it to another cpu. Taking rq->lock wrecks > that for sure.
| |