[lkml]   [2017]   [Sep]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [RFC] sched/fair: Use wake_q length as a hint for wake_wide
> On Tue, Sep 19, 2017 at 3:05 AM, Brendan Jackman
> <> wrote:
>> On Mon, Sep 18 2017 at 22:15, Joel Fernandes wrote:
>>>> IIUC, if wake_affine() behaves correctly this trick wouldn't be
>>>> necessary on SMP systems, so it might be best guarded by the presence
>>> Actually wake_affine doesn't check for balance if previous/next cpu
>>> are within the same shared cache domain. The difference is some time
>>> ago it would return true for shared cache but now it returns false as
>>> of 4.14-rc1:
>>> Since it would return false in the above wake up cases for task 1 and
>>> 2, it would then run select_idle_sibling on the previous CPU which is
>>> also within the big cluster, so I don't think it will make a
>>> difference in this case... Infact what it returns probably doesn't
>>> matter.
>> So my paragraph here was making a leap in reasoning, let me try to fill
>> the gap: On SMP these tasks never need to move around. If by some chance
>> they did get coscheduled, the first load balance would spread them out and
>> then every time they wake up from then on, prev_cpu is the sensible
>> choice. So it will look something like:
>> v CPU v ->time->
>> -------------
>> { 0 (SAME) 11111111111
>> cache { -------------
>> { 1 (SAME) 222222222222|
>> -------------
>> { 2 (SAME) 33333333333
>> cache { -------------
>> { 3 (SAME) 44444444444
>> -------------
>> So here, task 2 wakes up the other guys and when it's doing tasks 3 and
>> 4, prev_cpu and smp_processor_id() don't share a cache, so IIUC its'
>> basically wake_affine's job to decide between prev_cpu and
>> smp_processor_id(). So "if wake_affine is behaving correctly" the
>> problem that this patch aims to solve (i.e. the fact that we overload
>> the waker's LLC domain because of bias towards prev_cpu) does not arise
>> on SMP.
> Yes SMP, but your patch is for solving a problem for non-SMP. So your
> original statement about wake_affine solving any problem for SMP is
> not relevant I feel :-P. I guess you can just kill this para from the
> commit message to prevent confusion.

Ok I take that back, you were talking about guarding this feature by

I don't think that protection would be helpful because you can have
the same issue if the tasks do different amount of work on SMP. So in
that case some threads might still complete before the others and you
run into the same thing.


- Joel

 \ /
  Last update: 2017-09-20 07:06    [W:0.061 / U:9.296 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site