Re: [PATCH] sched/fair: Rate limit calls to update_blocked_averages() for NOHZ

[Dietmar Eggemann]

On 25/01/2021 18:30, Vincent Guittot wrote:
> On Mon, 25 Jan 2021 at 11:45, Dietmar Eggemann <dietmar.eggemann@xxxxxxx> wrote:
>>
>> On 22/01/2021 20:10, Joel Fernandes wrote:
>>> Hi Vincent,
>>>
>>> Thanks for reply. Please see the replies below:
>>>
>>> On Fri, Jan 22, 2021 at 05:56:22PM +0100, Vincent Guittot wrote:
>>>> On Fri, 22 Jan 2021 at 16:46, Joel Fernandes (Google)
>>>> <joel@xxxxxxxxxxxxxxxxx> wrote:

[...]

>> If I understood you correctly, you want to avoid these frequent calls
>> to update_blocked_averages() here to further avoid invoking sched_util
>> via update_blocked_averages() -> cpufreq_update_util() (since 'decayed'
>> is set) very often in your setup.
> 
> So It's not clear if the problem that joel wants to raise, is about:
> - the running time of  update_blocked_averages
> - the running time of the cpufreq_update_util which is called because
> utilization has decayed during the update of blocked load
> - the wake up latency because of newly_idle lb

Pretty much so.

IIRC his interest is driven by the fact that he saw much less activity
in newly_idle lb and therefore cpufreq_update_util on a system with the
same kernel and userspace but with less CPUs (i.e. also smaller
frequency domains) and less cgroups (with blocked load) and started
wondering why.

I assume that since he understands this environment now much better, he
should be able to come up with better test numbers to show if there is a
performance issue on his 2+6 DynamIQ system and if yes, where exactly in
this code path.