[PATCH 0/2] Fix imbalance issue when balancing fork

[Adam Li]

Load imbalance is observed when the workload frequently forks new threads.
Due to CPU affinity, the workload can run on CPU 0-7 in the first
group, and only on CPU 8-11 in the second group. CPU 12-15 are always idle.

{ 0 1 2 3 4 5 6 7 } {8 9 10 11 12 13 14 15}
  * * * * * * * *    * * *  *

When looking for dst group for newly forked threads, in many times
update_sg_wakeup_stats() reports the second group has more idle CPUs
than the first group. The scheduler thinks the second group is less
busy. Then it selects least busy CPUs among CPU 8-11. So CPU 8-11 can be
crowded with newly forked threads, at the same time CPU 0-7 can be idle.

The first patch 'Only update stats of allowed CPUs when looking for dst
group' *alone* can fix this imbalance issue.

And I think the second patch also makes sense in this scenario. If group
weight includes CPUs a task cannot use, group classification can be
incorrect. Please comment.

Adam Li (2):
  sched/fair: Only update stats of allowed CPUs when looking for dst
    group
  sched/fair: Only count group weight for allowed CPUs when looking for
    dst group

 kernel/sched/fair.c | 6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

-- 
2.34.1