On 23/11/2020 14:42, Vincent Guittot wrote: > On Mon, 23 Nov 2020 at 12:27, Dietmar Eggemann <dietmar.eggem...@arm.com> > wrote: >> >> On 23/11/2020 11:05, Vincent Guittot wrote: >>> On Mon, 23 Nov 2020 at 10:30, Dietmar Eggemann <dietmar.eggem...@arm.com> >>> wrote: >>>> >>>> On 20/11/2020 09:56, Peter Zijlstra wrote: >>>>> On Fri, Nov 20, 2020 at 08:55:27AM +0100, Peter Zijlstra wrote:
[...] >> I thought the question was whether 'runnable_avg = 1.5 x >> SCHED_CAPACITY_SCALE' is a good threshold to decide to drive frequency >> by runnable_avg or util_avg. > > we can't use SCHED_CAPACITY_SCALE and must use cpu's capacity To get some idle time on the LITTLE CPU I extended the time in which task3 and then task2-3 run on CPU5 to 128ms. The moment CPU4 has some idle time for the first time after the first migration (~207ms), the runnable_avg drops from 1323 (task0: 316, task1: 1020) to 1. I can't see the dependency to the CPU capacity here. Util_avg is also larger than the CPU capacity. The steep fall from runnable=1323 to 1 is due to lost_idle_time update. ... migration/4-29 [004] 60.000034: sched_migrate_task: comm=task3-3 pid=1690 prio=101 orig_cpu=4 dest_cpu=5 migration/4-29 [004] 60.000046: sched_pelt_cfs: cpu=4 path=/ load=163618 runnable=2296 util=748 ... migration/4-29 [004] 60.142088: sched_migrate_task: comm=task2-2 pid=1689 prio=101 orig_cpu=4 dest_cpu=5 migration/4-29 [004] 60.142100: sched_pelt_cfs: cpu=4 path=/ load=93358 runnable=1325 util=628 ... task0-0-1687 [004] 60.201385: sched_pelt_se: cpu=4 path=(null) comm=task0-0 pid=1687 load=22317 runnable=316 util=316 task0-0-1687 [004] 60.201387: sched_pelt_cfs: cpu=4 path=/ load=93978 runnable=1336 util=788 ... task1-1-1688 [004] 60.207225: sched_pelt_se: cpu=4 path=(null) comm=task1-1 pid=1688 load=71610 runnable=1020 util=497 task1-1-1688 [004] 60.207227: sched_pelt_cfs: cpu=4 path=/ load=93017 runnable=1323 util=800 <idle>-0 [004] 60.207254: cpu_idle: state=0 cpu_id=4 ... <idle>-0 [004] 60.209397: sched_pelt_cfs: cpu=4 path=/ load=80 runnable=1 util=0 ...
