lkml.org 
[lkml]   [2014]   [Jul]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH 2/2] sched: Rewrite per entity runnable load average tracking
Date
Yuyang Du <yuyang.du@intel.com> writes:

> Thanks, Peter.
>
> On Wed, Jul 09, 2014 at 08:45:43PM +0200, Peter Zijlstra wrote:
>
>> 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).
>
> Strictly, migrating wakee task on remote CPU entails two steps:
>
> (1) Catch up with task's queue's last_update_time, and then substract
>
> (2) Cache up with "current" time of remote CPU (for comparable matter), and then
> on new CPU, change to the new timing source (when enqueue)
>
> So I will try sched_clock_cpu(remote_cpu) for step (2). For step (2), maybe we
> should not use cfs_rq_clock_task anyway, since the task is about to going
> to another CPU/queue. Is this right?

So, sched_clock(_cpu) can be arbitrarily far off of cfs_rq_clock_task, so you
can't really do that. Ideally, yes, you would account for any time since
the last update and account that time as !runnable. However, I don't
think there is any good way to do that, and the current code doesn't.

>
> I made another mistake. Should not only track task entity load, group entity
> (as an entity) is also needed. Otherwise, task_h_load can't be done correctly...
> Sorry for the messup. But this won't make much change in the codes.

This will increase it to 2x __update_load_avg per cgroup per
enqueue/dequeue. What does this (and this patch in general) do to
context switch cost at cgroup depth 1/2/3?

>
> Thanks,
> Yuyang
>
>> 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?
>>
>> 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.


\
 
 \ /
  Last update: 2014-07-10 20:01    [W:0.165 / U:8.320 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site