Re: [v4.8-rc1 Regression] sched/fair: Apply more PELT fixes

[Morten Rasmussen]

On Tue, Oct 18, 2016 at 01:56:51PM +0200, Vincent Guittot wrote:
> Le Tuesday 18 Oct 2016 à 12:34:12 (+0200), Peter Zijlstra a écrit :
> > On Tue, Oct 18, 2016 at 11:45:48AM +0200, Vincent Guittot wrote:
> > > On 18 October 2016 at 11:07, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > > > So aside from funny BIOSes, this should also show up when creating
> > > > cgroups when you have offlined a few CPUs, which is far more common I'd
> > > > think.
> > > 
> > > The problem is also that the load of the tg->se[cpu] that represents
> > > the tg->cfs_rq[cpu] is initialized to 1024 in:
> > > alloc_fair_sched_group
> > >      for_each_possible_cpu(i) {
> > >          init_entity_runnable_average(se);
> > >             sa->load_avg = scale_load_down(se->load.weight);
> > > 
> > > Initializing  sa->load_avg to 1024 for a newly created task makes
> > > sense as we don't know yet what will be its real load but i'm not sure
> > > that we have to do the same for se that represents a task group. This
> > > load should be initialized to 0 and it will increase when task will be
> > > moved/attached into task group
> > 
> > Yes, I think that makes sense, not sure how horrible that is with the
> 
> That should not be that bad because this initial value is only useful for
> the few dozens of ms that follow the creation of the task group

IMHO, it doesn't make much sense to initialize empty containers, which
group sched_entities really are, to 1024. It is meant to represent what
is in it, and a creation it is empty, so in my opinion initializing it
to zero make sense.
 
> > current state of things, but after your propagate patch, that
> > reinstates the interactivity hack that should work for sure.

It actually works on mainline/tip as well.

As I see it, the fundamental problem is keeping group entities up to
date. Because the load_weight and hence se->avg.load_avg each per-cpu
group sched_entity depends on the group cfs_rq->tg_load_avg_contrib for
all cpus (tg->load_avg), including those that might be empty and
therefore not enqueued, we must ensure that they are updated some other
way. Most naturally as part of update_blocked_averages().

To guarantee that, it basically boils down to making sure:
Any cfs_rq with a non-zero tg_load_avg_contrib must be on the
leaf_cfs_rq_list.

We can do that in different ways: 1) Add all cfs_rqs to the
leaf_cfs_rq_list at task group creation, or 2) initialize group
sched_entity contributions to zero and make sure that they are added to
leaf_cfs_rq_list as soon as a sched_entity (task or group) is enqueued
on it.

Vincent patch below gives us the second option.

>  kernel/sched/fair.c | 9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 8b03fb5..89776ac 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -690,7 +690,14 @@ void init_entity_runnable_average(struct sched_entity *se)
>  	 * will definitely be update (after enqueue).
>  	 */
>  	sa->period_contrib = 1023;
> -	sa->load_avg = scale_load_down(se->load.weight);
> +	/*
> +	 * Tasks are intialized with full load to be seen as heavy task until
> +	 * they get a chance to stabilize to their real load level.
> +	 * group entity are intialized with null load to reflect the fact that
> +	 * nothing has been attached yet to the task group.
> +	 */
> +	if (entity_is_task(se))
> +		sa->load_avg = scale_load_down(se->load.weight);
>  	sa->load_sum = sa->load_avg * LOAD_AVG_MAX;
>  	/*
>  	 * At this point, util_avg won't be used in select_task_rq_fair anyway

I would suggest adding a comment somewhere stating that we need to keep
group cfs_rqs up to date:

-----
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index abb3763dff69..2b820d489be0 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6641,6 +6641,11 @@ static void update_blocked_averages(int cpu)
 		if (throttled_hierarchy(cfs_rq))
 			continue;
 
+		/*
+		 * Note that _any_ leaf cfs_rq with a non-zero tg_load_avg_contrib
+		 * _must_ be on the leaf_cfs_rq_list to ensure that group shares
+		 * are updated correctly.
+		 */
 		if (update_cfs_rq_load_avg(cfs_rq_clock_task(cfs_rq), cfs_rq, true))
 			update_tg_load_avg(cfs_rq, 0);
 	}
-----

I did a couple of simple tests on tip/sched/core to test whether
Vincent's fix works even without reflecting group load/util in the group
hierarchy:

Juno (2xA57+4xA53)

tip:
	grouped hog(1) alone: 2841
	non-grouped hogs(6) alone: 40830
	grouped hog(1): 218
	non-grouped hogs(6): 40580

tip+vg:
	grouped hog alone: 2849
	non-grouped hogs(6) alone: 40831
	grouped hog: 2363
	non-grouped hogs: 38418

See script below for details, but we basically see that the grouped task
is not getting its 'fair' share on tip, while it does with Vincent's
patch.

To summarize, I think Vincent's patch makes sense and works :-) More
testing is needed of cause to see if there are other problems.

-----

# Create 100 task groups:
for i in `seq 1 100`;
do
        cgcreate -g cpu:/root/test$i
done

NCPUS=$(grep -c ^processor /proc/cpuinfo)

# Run single cpu hog inside task group on first cpu _alone_:
cgexec -g cpu:/root/test100 taskset 0x01 sysbench --test=cpu \
--num-threads=1 --max-time=5 --max-requests=1000000 run | \
awk '{if ($4=="events:") {print "grouped hog(1) alone: " $5}}'

# Run cpu hogs outside task group _alone_:
sysbench --test=cpu --num-threads=$NCPUS --max-time=10 \
--max-requests=1000000 run | awk '{if ($4=="events:") \
{print "non-grouped hogs('$NCPUS') alone: " $5}}'

# Run cpu hogs outside task group:
sysbench --test=cpu --num-threads=$NCPUS --max-time=10 \
--max-requests=1000000 run | awk '{if ($4=="events:") \
{print "non-grouped hogs('$NCPUS'): " $5}}' &

# Run single cpu hog inside task group on first cpu:
cgexec -g cpu:/root/test100 taskset 0x01 sysbench \
--test=cpu --num-threads=1 --max-time=5 \
--max-requests=1000000 run | awk '{if ($4=="events:") \
{print "grouped hog(1): " $5}}'

wait

# Delete task groups:
for i in `seq 1 100`;
do
        cgdelete -g cpu:/root/test$i
done