lkml.org 
[lkml]   [2017]   [Apr]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v2] sched/cfs: make util/load_avg more stable
From
Date
On 25/04/17 13:40, Vincent Guittot wrote:
> On 25 April 2017 at 13:05, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote:
>> On 19/04/17 17:54, Vincent Guittot wrote:
>>> In the current implementation of load/util_avg, we assume that the ongoing
>>> time segment has fully elapsed, and util/load_sum is divided by LOAD_AVG_MAX,
>>> even if part of the time segment still remains to run. As a consequence, this
>>> remaining part is considered as idle time and generates unexpected variations
>>> of util_avg of a busy CPU in the range ]1002..1024[ whereas util_avg should
>>
>> Why do you use the square brackets the other way around? Just curious.
>
> This refers to the very beginning and very end of time segment formulas below.
> That being said, 1024 is not reachable because at very end we have :
> 1024*MAX_LOAD_AVG*y+1024*1023 = 1023,9997
>
> 1002 is not reachable because at very beg we have
> 1024*MAX_LOAD_AVG*y+ 1024*0 = 1002,0577
>
> But we are working with integer so [1002..1024[ is probably more correct

OK, this is with y = 32nd-rt(0.5) exactly, understood.

I assume you mean LOAD_AVG_MAX instead of MAX_LOAD_AVG.

>> 1002 stands for 1024*y^1 w/ y = 4008/4096 or y^32 = 0.5, right ? Might
>> be worth mentioning.
>>
>>> stay at 1023.
>>>
>>> In order to keep the metric stable, we should not consider the ongoing time
>>> segment when computing load/util_avg but only the segments that have already
>>> fully elapsed. Bu to not consider the current time segment adds unwanted
>>> latency in the load/util_avg responsivness especially when the time is scaled
>>> instead of the contribution. Instead of waiting for the current time segment
>>> to have fully elapsed before accounting it in load/util_avg, we can already
>>> account the elapsed part but change the range used to compute load/util_avg
>>> accordingly.
>>>
>>> At the very beginning of a new time segment, the past segments have been
>>> decayed and the max value is MAX_LOAD_AVG*y. At the very end of the current
>>> time segment, the max value becomes 1024(us) + MAX_LOAD_AVG*y which is equal
>>> to MAX_LOAD_AVG. In fact, the max value is
>>> sa->period_contrib + MAX_LOAD_AVG*y at any time in the time segment.

s/MAX_LOAD_AVG/LOAD_AVG_MAX

>>>
>>> Taking advantage of the fact that MAX_LOAD_AVG*y == MAX_LOAD_AVG-1024, the
>>> range becomes [0..MAX_LOAD_AVG-1024+sa->period_contrib].
>>>
>>> As the elapsed part is already accounted in load/util_sum, we update the max
>>> value according to the current position in the time segment instead of
>>> removing its contribution.
>>
>> Removing its contribution stands for '- 1024' of 'LOAD_AVG_MAX - 1024'
>> which was added in patch 1/2?
>
> removing its contribution refers to "- sa->period_contrib * weight"
> and "- (running * sa->period_contrib << SCHED_CAPACITY_SHIFT))" in
> patch 1/2 of the previous version

Yup, makes sense, so the '-1024' is the influence of the current 'time
segment' (n = 0) then.

IMHO, the removing of contribution in patch 1/2 wouldn't take freq and
cpu scaling of contribution (which is still in accumulate_sum()) into
consideration.

>>> Suggested-by: Peter Zijlstra <peterz@infradead.org>
>>> Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org>
>>> ---
>>>
>>> Fold both patches in one
>>>
>>> kernel/sched/fair.c | 6 +++---
>>> 1 file changed, 3 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>> index 3f83a35..c3b8f0f 100644
>>> --- a/kernel/sched/fair.c
>>> +++ b/kernel/sched/fair.c
>>> @@ -3017,12 +3017,12 @@ ___update_load_avg(u64 now, int cpu, struct sched_avg *sa,
>>> /*
>>> * Step 2: update *_avg.
>>> */
>>> - sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX);
>>> + sa->load_avg = div_u64(sa->load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
>>> if (cfs_rq) {
>>> cfs_rq->runnable_load_avg =
>>> - div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX);
>>> + div_u64(cfs_rq->runnable_load_sum, LOAD_AVG_MAX - 1024 + sa->period_contrib);
>>> }
>>> - sa->util_avg = sa->util_sum / LOAD_AVG_MAX;
>>> + sa->util_avg = sa->util_sum / (LOAD_AVG_MAX - 1024 + sa->period_contrib);
>>>
>>> return 1;
>>> }
>>>

\
 
 \ /
  Last update: 2017-04-25 16:53    [W:0.087 / U:0.592 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site