Re: [RFT][PATCH v7 6/8] sched: idle: Select idle state before stopping the tick

[Thomas Ilsche <thomas.ilsche@tu-dresden.de>]

On 2018-03-28 10:13, Rafael J. Wysocki wrote:
> On Wed, Mar 28, 2018 at 12:10 AM, Rafael J. Wysocki <rjw@rjwysocki.net> wrote:
>> On Tuesday, March 27, 2018 11:50:02 PM CEST Thomas Ilsche wrote:
>>> On 2018-03-20 16:45, Rafael J. Wysocki wrote:
>>>> From: Rafael J. Wysocki <rafael.j.wysocki@intel.com>
>>>>
>>>> In order to address the issue with short idle duration predictions
>>>> by the idle governor after the tick has been stopped, reorder the
>>>> code in cpuidle_idle_call() so that the governor idle state selection
>>>> runs before tick_nohz_idle_go_idle() and use the "nohz" hint returned
>>>> by cpuidle_select() to decide whether or not to stop the tick.
>>>>
>>>> This isn't straightforward, because menu_select() invokes
>>>> tick_nohz_get_sleep_length() to get the time to the next timer
>>>> event and the number returned by the latter comes from
>>>> __tick_nohz_idle_enter().  Fortunately, however, it is possible
>>>> to compute that number without actually stopping the tick and with
>>>> the help of the existing code.
>>>
>>> I think something is wrong with the new tick_nohz_get_sleep_length.
>>> It seems to return a value that is too large, ignoring immanent
>>> non-sched timer.
>>
>> That's a very useful hint, let me have a look.
>>
>>> I tested idle-loop-v7.3. It looks very similar to my previous results
>>> on the first idle-loop-git-version [1]. Idle and traditional synthetic
>>> powernightmares are mostly good.
>>
>> OK
>>
>>> But it selects too deep C-states for short idle periods, which is bad
>>> for power consumption [2].
>>
>> That still needs to be improved, then.
>>
>>> I tracked this down with additional tests using
>>> __attribute__((optimize("O0"))) menu_select
>>> and perf probe. With this the behavior seems slightly different, but it
>>> shows that data->next_timer_us is:
>>> v4.16-rc6: the expected ~500 us [3]
>>> idle-loop-v7.3: many milliseconds to minutes [4].
>>> This leads to the governor to wrongly selecting C6.
>>>
>>> Checking with 372be9e and 6ea0577, I can confirm that the change is
>>> introduced by this patch.
>>
>> Yes, that's where the most intrusive reordering happens.
> 
> Overall, this is an interesting conundrum, because the case in
> question is when the tick should never be stopped at all during the
> workload and the code's behavior in that case should not change, so
> the change was not intentional.
> 
> Now, from walking through the code, as long as can_stop_idle_tick()
> returns 'true' all should be fine or at least I don't see why there is
> any difference in behavior in that case.
> 
> However, if can_stop_idle_tick() returns 'false' (for example, because
> need_resched() returns 'true' when it is evaluated), the behavior *is*
> different in a couple of ways.  I sort of know how that can be
> addressed, but I'd like to reproduce your results here.
> 
> Are you still using the same workload as before to trigger this behavior?
> 

Yes, the exact code I use is as follows

$ gcc poller.c -O3 -fopenmp -o poller_omp
$ GOMP_CPU_AFFINITY=0-35 ./poller_omp 500

#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>

int main(int argc, char *argv[])
{
     int sleep_us = 10000;
     if (argc == 2) {
         sleep_us = atoi(argv[1]);
     }

     #pragma omp parallel
     {
         while (1) {
             usleep(sleep_us);
         }
     }
}