Tracing showed that the per-cpu scanning cost of __select_idle_core()
(~120ns) was significantly higher than that of __select_idle_cpu()
(~40ns).
This means that, even when reduced to the minimal scan, we're still 3x
more expensive than the simple search.
perf annotate suggested this was mostly due to cache-misses on the
additional cpumasks used.
However, we can mitigate this by only doing the more expensive search
when there is a good chance it is beneficial. After all, when there
are no idle cores to be had, there's no point in looking for any
(SMT>2 might want to try without this).
Clearing has_idle_cores early (without an exhaustive search) should be
fine because we're eager to set it when a core goes idle again.
FOLD
1: 0.568188455 seconds time elapsed ( +- 0.40% )
2: 0.643264625 seconds time elapsed ( +- 1.27% )
5: 2.385378263 seconds time elapsed ( +- 1.12% )
10: 3.808555491 seconds time elapsed ( +- 1.46% )
20: 6.431994272 seconds time elapsed ( +- 1.21% )
40: 9.423539507 seconds time elapsed ( +- 2.07% )
FOLD+
1: 0.554694881 seconds time elapsed ( +- 0.42% )
2: 0.632730119 seconds time elapsed ( +- 1.84% )
5: 2.230432464 seconds time elapsed ( +- 1.17% )
10: 3.549957778 seconds time elapsed ( +- 1.55% )
20: 6.118364255 seconds time elapsed ( +- 0.72% )
40: 9.515406550 seconds time elapsed ( +- 1.74% )
Signed-off-by: Peter Zijlstra (Intel) <pet...@infradead.org>
---
kernel/sched/fair.c | 5 ++++-
kernel/sched/features.h | 2 +-
2 files changed, 5 insertions(+), 2 deletions(-)
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6382,6 +6382,8 @@ static int __select_idle_core(struct tas
}
}
+ set_idle_cores(target, 0);
+
return best_cpu;
}
@@ -6477,7 +6479,8 @@ static int select_idle_cpu(struct task_s
time = local_clock();
#ifdef CONFIG_SCHED_SMT
- if (sched_feat(SIS_FOLD) && static_branch_likely(&sched_smt_present))
+ if (sched_feat(SIS_FOLD) && static_branch_likely(&sched_smt_present) &&
+ test_idle_cores(target, false))
cpu = __select_idle_core(p, sd, target, nr, &loops);
else
#endif
--- a/kernel/sched/features.h
+++ b/kernel/sched/features.h
@@ -60,7 +60,7 @@ SCHED_FEAT(SIS_PROP, true)
SCHED_FEAT(SIS_AGE, true)
SCHED_FEAT(SIS_ONCE, true)
-SCHED_FEAT(SIS_FOLD, false)
+SCHED_FEAT(SIS_FOLD, true)
/*
* Issue a WARN when we do multiple update_rq_clock() calls
