On 19/12/2018 15:20, Vincent Guittot wrote:
[...]
>> Oh yes, I never said it didn't work - I was doing some investigation on
>> the reason as to why we'd need this fix, because it's wasn't explicit from
>> the commit message.
>>
>> The rounding errors are countered by the +1, yes, but I'd rather remove
>> the errors altogether and go for the snippet I suggested in my previous
>> reply.
>
> except that you don't always want to migrate all group load.
> I prefer keeping current algorithm and fix it for now. Trying "new"
> thing can come in a 2nd step
We already set the imbalance as the whole group load, we just introduce
rounding errors inbetween. As I've already said:
in update_sg_lb_stats() we do:
sgs->avg_load = (sgs->group_load*SCHED_CAPACITY_SCALE) /
sgs->group_capacity;
and in check_asym_packing() we do:
env->imbalance = DIV_ROUND_CLOSEST(
sds->busiest_stat.avg_load * sds->busiest_stat.group_capacity,
SCHED_CAPACITY_SCALE)
So we end up with something like:
group_load * SCHED_CAPACITY_SCALE * group_capacity
imbalance = --------------------------------------------------
group_capacity * SCHED_CAPACITY_SCALE
Which we could reduce down to:
imbalance = group_load
and not get any rounding errors.
