Re: Kernel 4.7rc3 - Performance drop 30-40% for SPECjbb2005 and SPECjvm2008 benchmarks against 4.6 kernel

From: Vincent Guittot
Date: Fri Jun 24 2016 - 09:24:04 EST


On 24 June 2016 at 15:09, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Fri, Jun 24, 2016 at 02:44:07PM +0200, Vincent Guittot wrote:
>
>> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> > index 22d64b3f5876..d4f6fb2f3057 100644
>> > --- a/kernel/sched/fair.c
>> > +++ b/kernel/sched/fair.c
>> > @@ -2484,7 +2484,7 @@ static inline long calc_tg_weight(struct task_group *tg, struct cfs_rq *cfs_rq)
>> > */
>> > tg_weight = atomic_long_read(&tg->load_avg);
>> > tg_weight -= cfs_rq->tg_load_avg_contrib;
>> > - tg_weight += cfs_rq->load.weight;
>> > + tg_weight += cfs_rq->avg.load_avg;
>>
>> IIUC, you are reverting
>> commit fde7d22e01aa (sched/fair: Fix overly small weight for
>> interactive group entities)
>
> Ah!, I hadn't yet done a git-blame on this. Right you are, we should
> have put a comment there.
>
> So the problem here is that since commit:
>
> 2159197d6677 ("sched/core: Enable increased load resolution on 64-bit kernels")
>
> load.weight and avg.load_avg are in different metrics. Which completely
> wrecked things.
>
> The obvious alternative is using:
>
> scale_load_down(cfs_rq->load.weight);
>
> Let me go run that through the benchmark.

Yes, looks to be good alternative

>
>> I have one question regarding the use of cfs_rq->avg.load_avg
>> cfs_rq->tg_load_avg_contrib is the sampling of cfs_rq->avg.load_avg so
>> I'm curious to understand why you use cfs_rq->avg.load_avg instead of
>> keeping cfs_rq->tg_load_avg_contrib. Do you think that the sampling is
>> not accurate enough to prevent any significant difference between both
>> when we use tg->load_avg ?
>
> I'm not entirely sure I understand your question; is it to the existence
> of calc_tg_weight()? That is, why use calc_tg_weight() and not use
> tg->load_avg as is?

Yes

>
> It seemed like a simple and cheap way to increase accuracy, nothing more
> behind it until the commit you referred to.

Thanks for the clarification.
I thought that the difference should always be smaller than 1/64th of
the cfs_rq->avg.load_avg thanks to update_tg_load_avg