sched/cfs: change initial value of runnable_avg
authorVincent Guittot <>
Wed, 24 Jun 2020 15:44:22 +0000 (17:44 +0200)
committerBorislav Petkov <>
Sun, 28 Jun 2020 15:01:20 +0000 (17:01 +0200)
Some performance regression on reaim benchmark have been raised with
  commit 070f5e860ee2 ("sched/fair: Take into account runnable_avg to classify group")

The problem comes from the init value of runnable_avg which is initialized
with max value. This can be a problem if the newly forked task is finally
a short task because the group of CPUs is wrongly set to overloaded and
tasks are pulled less agressively.

Set initial value of runnable_avg equals to util_avg to reflect that there
is no waiting time so far.

Fixes: 070f5e860ee2 ("sched/fair: Take into account runnable_avg to classify group")
Reported-by: kernel test robot <>
Signed-off-by: Vincent Guittot <>
Signed-off-by: Peter Zijlstra (Intel) <>

index cbcb2f71599bbb36c2818536a189c234d70827c5..658aa7a2ae6f7d75e11012f437a45edd8befcc71 100644 (file)
@@ -806,7 +806,7 @@ void post_init_entity_util_avg(struct task_struct *p)
-       sa->runnable_avg = cpu_scale;
+       sa->runnable_avg = sa->util_avg;
        if (p->sched_class != &fair_sched_class) {