Re: [RFC] Documentation/scheduler/schedutil.txt

[Dietmar Eggemann]

On 23/11/2020 11:05, Vincent Guittot wrote:
> On Mon, 23 Nov 2020 at 10:30, Dietmar Eggemann <dietmar.eggemann@xxxxxxx> wrote:
>>
>> On 20/11/2020 09:56, Peter Zijlstra wrote:
>>> On Fri, Nov 20, 2020 at 08:55:27AM +0100, Peter Zijlstra wrote:
>>>>  - In saturated scenarios task movement will cause some transient dips,
>>>>    suppose we have a CPU saturated with 4 tasks, then when we migrate a task
>>>>    to an idle CPU, the old CPU will have a 'running' value of 0.75 while the
>>>>    new CPU will gain 0.25. This is inevitable and time progression will
>>>>    correct this. XXX do we still guarantee f_max due to no idle-time?
>>>
>>> Do we want something like this? Is the 1.5 threshold sane? (it's been too
>>> long since I looked at actual numbers here)
>>
>> Did some tests on a big.little system:
>>
>>  (1) rt-app workload on big CPU:
>>
>>  - task0-3 (runtime/period=4000us/16000us, started with
>>    4000us delay to each other) run on CPU1
>>  - then task3 migrates to CPU2 and runs there for 64ms
>>  - then task2 migrates to CPU2 too and both tasks run there
>>    for another 64ms
>>
>> ...
>>     task3-3-1684  [001]  3982.798729: sched_pelt_cfs:       cpu=1 path=/ load=232890 runnable=3260 util=1011
>> migration/1-14    [001]  3982.798756: sched_migrate_task:   comm=task3-3 pid=1684 prio=101 orig_cpu=1 dest_cpu=2*
>> migration/1-14    [001]  3982.798767: sched_pelt_cfs:       cpu=1 path=/ load=161374 runnable=2263 util=*700* <-- util dip !!!
>>     task1-1-1682  [001]  3982.799802: sched_pelt_cfs:       cpu=1 path=/ load=160988 runnable=2257 util=706
>> ...
>>     task2-2-1683  [001]  3982.849123: sched_pelt_cfs:       cpu=1 path=/ load=161124 runnable=2284 util=904
>>     task2-2-1683  [001]  3982.851960: sched_pelt_cfs:       cpu=1 path=/ load=160130 runnable=2271 util=911
>> migration/1-14    [001]  3982.851984: sched_migrate_task:   comm=task2-2 pid=1683 prio=101 orig_cpu=1 dest_cpu=2**
>> migration/1-14    [001]  3982.851995: sched_pelt_cfs:       cpu=1 path=/ load=88672 runnable=*1257* util=512 <-- runnable below 1536
>>     task1-1-1682  [001]  3982.852983: sched_pelt_cfs:       cpu=1 path=/ load=88321 runnable=1252 util=521
>> ...
>>
>>
>> *  task1,2,3 remain on CPU1 and still have to catch up, no idle
>>    time on CPU1
>>
>> ** task 1,2 remain on CPU1, there is idle time on CPU1!
>>
>>
>> (2) rt-app workload on LITTLE CPU (orig cpu_capacity: 446)
>>
>>  - task0-3 (runtime/period=1742us/16000us, started with
>>    4000us delay to each other) run on CPU4
>>  - then task3 migrates to CPU5 and runs there for 64ms
>>  - then task2 migrates to CPU5 too and both tasks run there
>>    for another 64ms
>>
>> ...
>>     task1-1-1777  [004]   789.443015: sched_pelt_cfs:       cpu=4 path=/ load=234718 runnable=3018 util=976
>> migration/4-29    [004]   789.444718: sched_migrate_task:   comm=task3-3 pid=1779 prio=101 orig_cpu=4 dest_cpu=5*
>> migration/4-29    [004]   789.444739: sched_pelt_cfs:       cpu=4 path=/ load=163543 runnable=2114 util=*778* <--util dip !!!
>>     task2-2-1778  [004]   789.447013: sched_pelt_cfs:       cpu=4 path=/ load=163392 runnable=2120 util=777
>> ...
>>     task1-1-1777  [004]   789.507012: sched_pelt_cfs:       cpu=4 path=/ load=164482 runnable=2223 util=879
>> migration/4-29    [004]   789.508023: sched_migrate_task:   comm=task2-2 pid=1778 prio=101 orig_cpu=4 dest_cpu=5**
>> migration/4-29    [004]   789.508044: sched_pelt_cfs:       cpu=4 path=/ load=94099 runnable=*1264* util=611 <-- runnable below 1536
>>     task0-0-1776  [004]   789.511011: sched_pelt_cfs:       cpu=4 path=/ load=93898 runnable=1264 util=622
>> ...
>>
>> *  task1,2,3 remain on CPU1 and still have to catch up, no idle
>>    time on CPU1
>>
>> ** task 1,2 remain on CPU1, no idle time on CPU1 yet.
>>
>> So for the big CPU, there is idle time and for the LITTLE there
>> isn't with runnable below the threshold.
> 
> I'm not sure to catch what you want to highlight with your tests ?

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.

[...]