These patches mitigate latency caused by newidle_balance() on large systems when PREEMPT_RT is enabled, by enabling interrupts when the lock is dropped, and exiting early at various points if an RT task is runnable on the current CPU. On a system with 128 CPUs, these patches dropped latency (as measured by a 12 hour rteval run) from 1045us to 317us (when applied to 5.12.0-rc3-rt3). I tried a couple scheduler benchmarks (perf bench sched pipe, and sysbench threads) to try to determine whether the overhead is measurable on non-RT, but the results varied widely enough (with or without the patches) that I couldn't draw any conclusions from them. So at least for now, I limited the balance callback change to when PREEMPT_RT is enabled. Link to v1 RFC patches: https://lore.kernel.org/lkml/20200428050242.17717-1-swood@xxxxxxxxxx/ Scott Wood (3): sched/fair: Call newidle_balance() from balance_callback on PREEMPT_RT sched/fair: Enable interrupts when dropping lock in newidle_balance() sched/fair: break out of newidle balancing if an RT task appears kernel/sched/fair.c | 66 ++++++++++++++++++++++++++++++++++++++------ kernel/sched/sched.h | 6 ++++ 2 files changed, 64 insertions(+), 8 deletions(-) -- 2.27.0
