[lkml]   [2008]   [Dec]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [patch] sched: fix uneven per-cpu task_group share distribution

* Peter Zijlstra <> wrote:

> On Mon, 2008-12-15 at 23:37 -0800, Ken Chen wrote:
> > While testing CFS scheduler on linux-2.6-tip tree
> > git://
> >
> > We found that task which is pinned to a CPU could be starved relative to its
> > allocated fair share.
> >
> > The per-cpu sched_enetity load share calculation in tg_shares_up /
> > update_group_shares_cpu distributes total task_group's share among all CPUs
> > for a given SD domain, this would dilute task_group's per-cpu share because
> > it distributes share to CPU that even has no load. The trapped share is now
> > un-consumable and it leads to fair share starvation on the runnable CPU.
> > Peter was right that it is still required for the low level function to make
> > distinction between a boosted share that don't have any load and actual tg
> > share that should be distributed among CPUs in which the tg is running.
> >
> > Patch to add that boost and we think the scheduler should only boost one
> > times of tg shares over all empty CPU that don't have any load for the
> > specific task_group in order to bound maximum temporary boost that a given
> > task_group can have.
> Yeah I was worried about you removing that boost, but since you seemed
> to be doing serious testing on the thing.. A well, this sounds good.
> > Signed-off-by: Ken Chen <>
> OK, patch looks good too, thanks!
> Acked-by: Peter Zijlstra <>

applied to tip/cpus4096. [not in the tip/sched/core branch, because
cpumask_t changes conflict with Ken's fix so this is the proper ordering.]


 \ /
  Last update: 2008-12-18 14:45    [W:0.043 / U:0.080 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site