[lkml]   [2021]   [Aug]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [PATCH 8/9] sched/fair: select idle cpu from idle cpumask for task wakeup
On Mon, Aug 02, 2021 at 10:41:13AM +0000, Song Bao Hua (Barry Song) wrote:
> Hi Mel, Aubrey,
> A similar thing Yicong and me has discussed is having a mask or a count for
> idle cores. And then we can only scan idle cores in this mask in
> select_idle_cpu().

Either approach would require a lot of updates.

> A obvious problem is that has_idle_cores is a bool, it can seriously lag
> from the real status. I mean, after system enters the status without idle
> cores, has_idle_cores could be still true.


> Right now, we are setting has_idle_cores to true while cpu enters idle
> and its smt sibling is also idle. But we are setting has_idle_cores to
> false only after we scan all cores in a llc.
> So we have thought for a while to provide an idle core mask. But never
> really made a workable patch.
> Mel's patch7/9 limits the number of cores which will be scanned in
> select_idle_cpu(), it might somehow alleviate the problem we redundantly
> scan all cores while we actually have no idle core even has_idle_cores
> is true.

I prototyped a patch that tracked the location of a known idle core and
use it as a starting hint for a search. It's cleared if the core is
selected for use. Unfortunately, it was not a universal win so was
dropped for the moment but either way, improving the accurate of
has_idle_cores without being too expensive would be niuce.

> However, if we can get idle core mask, it could be also good to
> select_idle_core()? Maybe after that, we don't have to enforce
> proportional scan limits while scanning for an idle core?

To remove the limits entirely, it would have to be very accurate and
it's hard to see how that can be done without having a heavily updated
shared cache line.

Mel Gorman

 \ /
  Last update: 2021-08-04 12:26    [W:0.080 / U:0.032 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site