lkml.org 
[lkml]   [2022]   [Jan]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH] sched/fair: Prefer small idle cores for forkees
From

On 1/21/2022 3:47 PM, Vincent Donnefort wrote:
> On Thu, Jan 20, 2022 at 10:15:07PM +0530, Chitti Babu Theegala wrote:
>>
>>
>> On 1/13/2022 10:05 PM, Vincent Donnefort wrote:
>>> On Wed, Jan 12, 2022 at 08:09:02PM +0530, Chitti Babu Theegala wrote:
>>>> Newly forked threads don't have any useful utilization data yet and
>>>> it's not possible to forecast their impact on energy consumption.
>>>> update_pick_idlest These forkees (though very small, most times) end up waking big
>>>> cores from deep sleep for that very small durations.
>>>>
>>>> Bias all forkees to small cores to prevent waking big cores from deep
>>>> sleep to save power.
>>>
>>> This bias might be interesting for some workloads, but what about the
>>> others? (see find_energy_efficient_cpu() comment, which discusses forkees).
>>>
>>
>> Yes, I agree with the find_energy_efficient_cpu() comment that we don't have
>> any useful utilization data yet and hence not possible to forecast. However,
>> I don't see any point in penalizing the power by waking up bigger cores
>> which are in deep sleep state for very small workloads.
>>
>> This patch helps lighter workloads during idle conditions w.r.t power POV.
>> For active (interactive or heavier) workloads, on most big.Little systems'
>> these foreground tasks get pulled into gold affined cpu-sets where this
>> patch would not play any spoilsport. Even for systems with such cpu-sets not
>> defined, heavy workloads might need just another 1 or 2 scheduling windows
>> for ramping to better freq or core.
>
> Scheduling windows? I suppose you do not refer to PELT here, so I'm not sure
> this argument applies here.

Sorry. I didn’t mean it to be WALT. I meant that ramp up would happen in
next couple of ms which can give very small penalty for such heavy
workloads for the initial ms.

>
> Beside, CFS always bias toward performance (except feec(), which does it in a
> lesser extent).
>

Yes, aware that CFS is perf bias. Can we have a knob atleast which can
turn-on such power friendly features ?

>>
>>>>
>>>> Signed-off-by: Chitti Babu Theegala <quic_ctheegal@quicinc.com>
>>>> ---
>>>> kernel/sched/fair.c | 16 +++++++++++-----
>>>> 1 file changed, 11 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>>> index 6e476f6..d407bbc 100644
>>>> --- a/kernel/sched/fair.c
>>>> +++ b/kernel/sched/fair.c
>>>> @@ -5976,7 +5976,7 @@ static int wake_affine(struct sched_domain *sd, struct task_struct *p,
>>>> }
>>>> static struct sched_group *
>>>> -find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu);
>>>> +find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu, int sd_flag);
>>>> /*
>>>> * find_idlest_group_cpu - find the idlest CPU among the CPUs in the group.
>>>> @@ -6063,7 +6063,7 @@ static inline int find_idlest_cpu(struct sched_domain *sd, struct task_struct *p
>>>> continue;
>>>> }
>>>> - group = find_idlest_group(sd, p, cpu);
>>>> + group = find_idlest_group(sd, p, cpu, sd_flag);
>>>> if (!group) {
>>>> sd = sd->child;
>>>> continue;
>>>> @@ -8997,7 +8997,8 @@ static inline void update_sg_wakeup_stats(struct sched_domain *sd,
>>>> static bool update_pick_idlest(struct sched_group *idlest,
>>>> struct sg_lb_stats *idlest_sgs,
>>>> struct sched_group *group,
>>>> - struct sg_lb_stats *sgs)
>>>> + struct sg_lb_stats *sgs,
>>>> + int sd_flag)
>>>> {
>>>> if (sgs->group_type < idlest_sgs->group_type)
>>>> return true;
>>>> @@ -9034,6 +9035,11 @@ static bool update_pick_idlest(struct sched_group *idlest,
>>>> if (idlest_sgs->idle_cpus > sgs->idle_cpus)
>>>> return false;
>>>> + /* Select smaller cpu group for newly woken up forkees */
>>>> + if ((sd_flag & SD_BALANCE_FORK) && (idlest_sgs->idle_cpus &&
>>>> + !capacity_greater(idlest->sgc->max_capacity, group->sgc->max_capacity)))
>>>> + return false;
>>>> +
>>>
>>> Energy biased placement should probably be applied only when EAS is enabled.
>>>
>>> It's especially true here, if all CPUs have the same capacity, capacity_greater
>>> would be always false. So unless I missed something, we wouldn't let the group_util
>>> evaluation happen, would we?
>>
>> True. I am uploading new version patch with a EAS enablement check in place.
>>
>>>
>>> [...]

\
 
 \ /
  Last update: 2022-01-25 08:39    [W:0.056 / U:0.264 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site