Messages in this thread | | | From | Mel Gorman <> | Subject | [PATCH 0/4] Throttle select_idle_sibling when a target domain is overloaded | Date | Fri, 20 Mar 2020 15:12:41 +0000 |
| |
This is a follow-on from the CPU/NUMA load balancer reconcilation after I noticed that select_idle_sibling() was doing excessive work. It was originally part of a larger series that merged select_idle_core, select_idle_sibling and select_idle_cpu as a single pass. Unfortunately, fixes have invalidated the tests multiple times so this series covers only one part for now as the tests are extremely time-consuming.
tip/sched/core as of March 13th was used as the baseline with "sched/fair: fix condition of avg_load calculation" applied which was just picked up by tip at the time of writing.
Patches 1-2 add schedstats to track the efficiency of select_idle_sibling(). Ordinarily they are disabled and are only really of use to a kernel developer. However, I find them more practical to work with than perf.
Patch 3 is a trivial micro-optimisation that avoids clearing part of a cpumask if a core has been found.
Patch 4 tracks whether a domain appeared to be overloaded during select_idle_cpu() so that future scans can abort early if necessary. This reduces the number of runqueues that are scanned uselessly when a domain is overloaded.
include/linux/sched/topology.h | 1 + kernel/sched/debug.c | 6 +++ kernel/sched/fair.c | 103 +++++++++++++++++++++++++++++++++++------ kernel/sched/features.h | 3 ++ kernel/sched/sched.h | 8 ++++ kernel/sched/stats.c | 9 ++-- 6 files changed, 113 insertions(+), 17 deletions(-)
-- 2.16.4
| |