Re: [RFC PATCH] sched/fair: use max_spare_cap_cpu if it is more energy efficient

[Dietmar Eggemann <dietmar.eggemann@arm.com>]

On 22/10/2021 06:05, Xuewen Yan wrote:
> Hi Chunguang
> 
> brookxu <brookxu.cn@gmail.com> 于2021年10月21日周四 下午4:24写道：
>>
>> From: Chunguang Xu <brookxu@tencent.com>
>>
>> When debugging EAS, I found that if the task is migrated to
>> max_spare_cap_cpu, even if the power consumption of pd is lower,

The task p hasn't been migrated yet. `max_spare_cap_cpu` here is only a
potential candidate CPU to be selected for p.

>> we still put the task on prev_cpu. Maybe we should fix it.
>>
>> Signed-off-by: Chunguang Xu <brookxu@tencent.com>
>> ---
>>  kernel/sched/fair.c | 4 +++-
>>  1 file changed, 3 insertions(+), 1 deletion(-)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index ff69f245b939..2ae7e03de6d2 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -6867,8 +6867,10 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu)
>>                 /* Evaluate the energy impact of using max_spare_cap_cpu. */
>>                 if (max_spare_cap_cpu >= 0) {
>>                         cur_delta = compute_energy(p, max_spare_cap_cpu, pd);
>> -                       if (cur_delta < base_energy_pd)
> 
> this is aimed to prevent the cur_delta < 0, and usuallly, when the
> task was put on the max_spare_cpu, the cur_power should be bigger than
> base_pd_power,
> if the cur_power < base_pd_power, the cpu util may have changed, at
> this time, we should keep prev_cpu.
> 
> You can look at below discuss and patch:
> https://lore.kernel.org/all/20210429101948.31224-3-Pierre.Gondois@arm.com/
> https://lore.kernel.org/all/CAB8ipk_vgtg5d1obH36BYfNLZosbwr2k_U3xnAD4=H5uZt_M_g@mail.gmail.com/

That's correct. `prev_delta < base_energy_pd` or `cur_delta <
base_energy_pd` indicate the rare case that `compute_energy() { ->
cpu_util_next() -> cpu util }` returns a higher energy value for the
perf domain w/o the task p than w/ it.

`base_energy_pd` stands for the energy spend on the CPUs of the Perf
Domain (PD) w/o considering the task p (compute_energy(p, *-1*, pd)),
`dst_cpu == -1`.

If this happens to a candidate CPU (prev_cpu or a per-PD
max_spare_cap_cpu) we bail out and return target (i.e. prev_cpu) because
we can't compare the energy values (prev_delta and best_delta) later on
in this case.

[...]