lkml.org 
[lkml]   [2020]   [Apr]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH 1/4] sched/topology: Store root domain CPU capacity sum
From
Date
On 09.04.20 16:13, Vincent Guittot wrote:
> On Thu, 9 Apr 2020 at 15:50, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote:
>>
>> On 08.04.20 19:03, Vincent Guittot wrote:
>>> On Wed, 8 Apr 2020 at 18:31, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote:
>>>>
>>>> On 08.04.20 14:29, Vincent Guittot wrote:
>>>>> On Wed, 8 Apr 2020 at 11:50, Dietmar Eggemann <dietmar.eggemann@arm.com> wrote:

[...]

>> And it looks like that asym_cpu_capacity_level() [topology.c] would fail
>> if we would use capacity_orig_of() instead of arch_scale_cpu_capacity().
>
> Yes I agree. See below
>
>> post_init_entity_util_avg() [fair.c] and sugov_get_util()
>> [cpufreq_schedutil.c] would be temporarily off until
>> update_cpu_capacity() has updated cpu_rq(cpu)->cpu_capacity_orig.
>
> I think that we could even get rid of this update in
> update_cpu_capacity(). cpu_capacity_orig should be set while building
> the sched_domain topology because the topology itself is built based
> on this max cpu capacity with asym_cpu_capacity_level(). So changing
> the capacity without rebuilding the domain could break the
> sched_domain topology correctness.

True. rq->cpu_capacity_orig could be set early in build_sched_domains(),
before the call to asym_cpu_capacity_level() or within this function.

> And we can't really set cpu_capacity_orig earlier during the boot
> because the capacity of b.L is set late during the boot and a rebuild
> of the sched_domain topology is then triggered.
>
>>
>> compute_energy() [fair.c] is guarded by sched_energy_enabled() from
>> being used at startup.
>>
>> scale_rt_capacity() could be changed in case we call it after the
>> cpu_rq(cpu)->cpu_capacity_orig = arch_scale_cpu_capacity(cpu) in
>> update_cpu_capacity().
>
> With the removal of the update in update_cpu_capacity(), we don't have
> a problem anymore, isn't it ?

True.

>> The Energy Model (and CPUfreq cooling) code would need
>> capacity_orig_of() exported. arch_scale_cpu_capacity() currently is
>> exported via include/linux/sched/topology.h.
>
> Not sure that we need to export it outside scheduler, they can still
> use arch_scale_cpu_capacity()

OK, let's change this for the task scheduler only.

>> I guess Pelt and 'scale invariant Deadline bandwidth enforcement' should
>> continue using arch_scale_cpu_capacity() in sync with
>> arch_scale_freq_capacity().
>
> Why can't they use capacity_orig_of ?
> we keep using arch_scale_freq_capacity() because it's dynamic but we
> don't really need to keep using arch_scale_cpu_capacity()

OK, Pelt is task scheduler so it can be changed here as well.

I'm going to create a patch following these ideas.

[...]

\
 
 \ /
  Last update: 2020-04-14 11:21    [W:0.089 / U:0.712 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site