On 06/30/2016 12:49 AM, Morten Rasmussen wrote:
> On Thu, Jun 23, 2016 at 02:20:48PM -0700, Sai Gurrappadi wrote:
>> Hi Morten,
>>
>> On 06/22/2016 10:03 AM, Morten Rasmussen wrote:
>>
>> [...]
>>
>>>
>>> +/*
>>> + * group_smaller_cpu_capacity: Returns true if sched_group sg has smaller
>>> + * per-cpu capacity than sched_group ref.
>>> + */
>>> +static inline bool
>>> +group_smaller_cpu_capacity(struct sched_group *sg, struct sched_group *ref)
>>> +{
>>> + return sg->sgc->max_capacity * capacity_margin <
>>> + ref->sgc->max_capacity * 1024;
>>> +}
>>> +
>>> static inline enum
>>> group_type group_classify(struct sched_group *group,
>>> struct sg_lb_stats *sgs)
>>> @@ -6892,6 +6903,19 @@ static bool update_sd_pick_busiest(struct lb_env
>>> *env,
>>> if (sgs->avg_load <= busiest->avg_load)
>>> return false;
>>>
>>> + if (!(env->sd->flags & SD_ASYM_CPUCAPACITY))
>>> + goto asym_packing;
>>> +
>>> + /* Candidate sg has no more than one task per cpu and has
>>> + * higher per-cpu capacity. Migrating tasks to less capable
>>> + * cpus may harm throughput. Maximize throughput,
>>> + * power/energy consequences are not considered.
>>> + */
>>> + if (sgs->sum_nr_running <= sgs->group_weight &&
>>> + group_smaller_cpu_capacity(sds->local, sg))
>>> + return false;
>>> +
>>> +asym_packing:
>>
>> What about the case where IRQ/RT work reduces the capacity of some of
>> these bigger CPUs? sgc->max_capacity might not necessarily capture
>> that case.
>
> Right, we could possibly improve this by using min_capacity instead, but
> we could end up allowing tasks to be pulled to lower capacity cpus just
> because one big cpu has reduced capacity due to RT/IRQ pressure and
> therefore has lowered the groups min_capacity.
>
> Ideally we should check all the capacities, but that complicates things
> a lot.
>
> Would you prefer min_capacity instead, or attempts to consider all the
> cpu capacities available in both groups?
>
min_capacity as a start works I think given that we are only trying to make
existing LB better, not necessarily optimizing for every case. Might have to
revisit this anyways for thermals etc.
Thanks,
-Sai
