Skip to content

Reduce migrations and conflicts with automatic NUMA balancing v2

Mel Gorman lkml 1 version In review

Get series with b4
b4 am [email protected]

Versions newest first

  1. v1 Latest6 patches 9 repliesIn review

Changelog since v1 from the cover letter

o Replace patch 4 with Peter's version
o Correct comments
o Add data on NAS OMP
o Rebase to 4.16-rc1

Patches in v1 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/6sched/fair: Avoid an unnecessary lookup of current CPU ID during wake_affinekernel/sched/fair.c+2 -30 replies
2/6sched/fair: Defer calculation of prev_eff_load in wake_affine until neededkernel/sched/fair.c+1 -10 replies
3/6sched/fair: Do not migrate on wake_affine_weight if weights are equalkernel/sched/fair.c+10 -10 replies
4/6sched/fair: Do not migrate due to a sync wakeup on exitkernel/sched/fair.c+1 -10 replies
5/6sched/fair: Consider SD_NUMA when selecting the most idle group to schedule onkernel/sched/fair.c+12 -00 replies
6/6sched/numa: Delay retrying placement for automatic NUMA balance after wake_affinekernel/sched/fair.c+56 -19 replies

Reviews trailers per patch

No Reviewed-by, Acked-by or Tested-by trailers for v1 yet.

Read the v1 thread

Cover letter [PATCH 0/6] Reduce migrations and conflicts with automatic NUMA balancing v2

This is a combination of "Reduce migrations due to load imbalance and process exits" and "Stop wake_affine fighting with automatic NUMA balancing" series.

There was a request for data on NAS OMP and how it compares with hard-binding of threads to CPUs that is included in the changelog for the relevant patch. I did not update most of the comparisons as a partial rerun of the workloads shows similar results.

Changelog since v1
o Replace patch 4 with Peter's version
o Correct comments
o Add data on NAS OMP
o Rebase to 4.16-rc1

The following series was motivated by the fact that higher migrations were observed than expected. The first two patches are minor cleanups. The third patch avoids migrations on wakeup when two CPUs are equally loaded which is particularly important in the case where the measured load is 0. The fourth patch avoids cross-node migrations of a parent process when the child exits. The fifth path resists new children being woken for the first time on a remote node but it's not barred completely as that may lead to saturation of a single node. The final patch forces automatic NUMA balancing to backoff if it disagrees with wake_affine about where a task should be running.

 kernel/sched/fair.c | 89 ++++++++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 82 insertions(+), 7 deletions(-)
Signature · 1 line
2.15.1