[PATCH] sched: fix incorrect PELT values on SMT

[PATCH] sched: fix incorrect PELT values on SMT

[Steve Muckle]

PELT scales its util_sum and util_avg values via
arch_scale_cpu_capacity(). If that function is passed the CPU's sched
domain then it will reduce the scaling capacity if SD_SHARE_CPUCAPACITY
is set. PELT does not pass in the sd however. The other caller of
arch_scale_cpu_capacity, update_cpu_capacity(), does. This means
util_sum and util_avg scale beyond the CPU capacity on SMT.

On an Intel i7-3630QM for example rq->cpu_capacity_orig is 589 but
util_avg scales up to 1024.

Fix this by passing in the sd in __update_load_avg() as well.

Signed-off-by: Steve Muckle <smuckle@linaro.org>
---
 kernel/sched/fair.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 61d485421bed..95d34b337152 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -2731,7 +2731,7 @@ __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
 	sa->last_update_time = now;
 
 	scale_freq = arch_scale_freq_capacity(NULL, cpu);
-	scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
+	scale_cpu = arch_scale_cpu_capacity(cpu_rq(cpu)->sd, cpu);
 
 	/* delta_w is the amount already accumulated against our next period */
 	delta_w = sa->period_contrib;
-- 
2.7.3

Re: [PATCH] sched: fix incorrect PELT values on SMT

[Wanpeng Li]

2016-08-19 9:55 GMT+08:00 Steve Muckle <steve.muckle@linaro.org>:
> PELT scales its util_sum and util_avg values via
> arch_scale_cpu_capacity(). If that function is passed the CPU's sched
> domain then it will reduce the scaling capacity if SD_SHARE_CPUCAPACITY
> is set. PELT does not pass in the sd however. The other caller of
> arch_scale_cpu_capacity, update_cpu_capacity(), does. This means
> util_sum and util_avg scale beyond the CPU capacity on SMT.
>
> On an Intel i7-3630QM for example rq->cpu_capacity_orig is 589 but
> util_avg scales up to 1024.
>
> Fix this by passing in the sd in __update_load_avg() as well.

I believe we notice this at least several months ago.
https://lkml.org/lkml/2016/5/25/228

>
> Signed-off-by: Steve Muckle <smuckle@linaro.org>
> ---
>  kernel/sched/fair.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 61d485421bed..95d34b337152 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -2731,7 +2731,7 @@ __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
>         sa->last_update_time = now;
>
>         scale_freq = arch_scale_freq_capacity(NULL, cpu);
> -       scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
> +       scale_cpu = arch_scale_cpu_capacity(cpu_rq(cpu)->sd, cpu);
>
>         /* delta_w is the amount already accumulated against our next period */
>         delta_w = sa->period_contrib;
> --
> 2.7.3
>



-- 
Regards,
Wanpeng Li

Re: [PATCH] sched: fix incorrect PELT values on SMT

[Steve Muckle]

On Fri, Aug 19, 2016 at 10:30:36AM +0800, Wanpeng Li wrote:
> 2016-08-19 9:55 GMT+08:00 Steve Muckle <steve.muckle@linaro.org>:
> > PELT scales its util_sum and util_avg values via
> > arch_scale_cpu_capacity(). If that function is passed the CPU's sched
> > domain then it will reduce the scaling capacity if SD_SHARE_CPUCAPACITY
> > is set. PELT does not pass in the sd however. The other caller of
> > arch_scale_cpu_capacity, update_cpu_capacity(), does. This means
> > util_sum and util_avg scale beyond the CPU capacity on SMT.
> >
> > On an Intel i7-3630QM for example rq->cpu_capacity_orig is 589 but
> > util_avg scales up to 1024.
> >
> > Fix this by passing in the sd in __update_load_avg() as well.
> 
> I believe we notice this at least several months ago.
> https://lkml.org/lkml/2016/5/25/228

Glad to see I'm not alone in thinking this is an issue.

It causes an issue with schedutil, effectively doubling the apparent
demand on SMT. I don't know the load balance code well enough offhand to
say whether it's an issue there.

cheers,
Steve

Re: [PATCH] sched: fix incorrect PELT values on SMT

[Dietmar Eggemann]

Hi Steve,

On 19/08/16 02:55, Steve Muckle wrote:
> PELT scales its util_sum and util_avg values via
> arch_scale_cpu_capacity(). If that function is passed the CPU's sched
> domain then it will reduce the scaling capacity if SD_SHARE_CPUCAPACITY
> is set. PELT does not pass in the sd however. The other caller of
> arch_scale_cpu_capacity, update_cpu_capacity(), does. This means
> util_sum and util_avg scale beyond the CPU capacity on SMT.
> 
> On an Intel i7-3630QM for example rq->cpu_capacity_orig is 589 but
> util_avg scales up to 1024.
> 
> Fix this by passing in the sd in __update_load_avg() as well.
> 
> Signed-off-by: Steve Muckle <smuckle@linaro.org>
> ---
>  kernel/sched/fair.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 61d485421bed..95d34b337152 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -2731,7 +2731,7 @@ __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
>  	sa->last_update_time = now;
>  
>  	scale_freq = arch_scale_freq_capacity(NULL, cpu);
> -	scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
> +	scale_cpu = arch_scale_cpu_capacity(cpu_rq(cpu)->sd, cpu);

Wouldn't you have to subscribe to this rcu pointer rq->sd w/ something
like 'rcu_dereference(cpu_rq(cpu)->sd)'?

IMHO, __update_load_avg() is called outside existing RCU read-side
critical sections as well so there would be a pair of
rcu_read_lock()/rcu_read_unlock() required in this case.

>  
>  	/* delta_w is the amount already accumulated against our next period */
>  	delta_w = sa->period_contrib;
> 

Re: [PATCH] sched: fix incorrect PELT values on SMT

[Morten Rasmussen]

Hi Steve,

On Thu, Aug 18, 2016 at 06:55:41PM -0700, Steve Muckle wrote:
> PELT scales its util_sum and util_avg values via
> arch_scale_cpu_capacity(). If that function is passed the CPU's sched
> domain then it will reduce the scaling capacity if SD_SHARE_CPUCAPACITY
> is set. PELT does not pass in the sd however. The other caller of
> arch_scale_cpu_capacity, update_cpu_capacity(), does. This means
> util_sum and util_avg scale beyond the CPU capacity on SMT.
> 
> On an Intel i7-3630QM for example rq->cpu_capacity_orig is 589 but
> util_avg scales up to 1024.

I can't convince myself whether this is the right thing to do. SMT is a
bit 'special' and it depends on how you model SMT capacity.

I'm no SMT expert, but the way I understand the current SMT capacity
model is that capacity_orig represents the capacity of the SMT-thread
when all its thread-siblings are busy. The true capacity of an
SMT-thread where all thread-siblings are idle is actually 1024, but we
don't model this (it would be nightmare to track when the capacity
should change). The capacity of a core with two or more SMT-threads is
chosen to be 1024 + smt_gain, where smt_gain is supposed represent the
additional throughput we gain for the additional SMT-threads. The reason
why we don't have 1024 per thread is that we would prefer to have only
one task per core if possible.

With util_avg scaling to 1024 a core (capacity = 2*589) would be nearly
'full' with just one always-running task. If we change util_avg to max
out at 589, it would take two always-running tasks for the combined
utilization to match the core capacity. So we may loose some bias
towards spreading for SMT systems.

AFAICT, group_is_overloaded() and group_has_capacity() would both be
affected by this patch.

Interestingly, Vincent recently proposed to set the SMT-thread capacity
to 1024 which would affectively make all the current SMT code redundant.
It would make things a lot simpler, but I'm not sure if we can get away
with it. It would need discussion at least.

Opinions?

Morten

Re: [PATCH] sched: fix incorrect PELT values on SMT

[Steve Muckle]

On Fri, Aug 19, 2016 at 04:30:39PM +0100, Morten Rasmussen wrote:
> Hi Steve,
> 
> On Thu, Aug 18, 2016 at 06:55:41PM -0700, Steve Muckle wrote:
> > PELT scales its util_sum and util_avg values via
> > arch_scale_cpu_capacity(). If that function is passed the CPU's sched
> > domain then it will reduce the scaling capacity if SD_SHARE_CPUCAPACITY
> > is set. PELT does not pass in the sd however. The other caller of
> > arch_scale_cpu_capacity, update_cpu_capacity(), does. This means
> > util_sum and util_avg scale beyond the CPU capacity on SMT.
> > 
> > On an Intel i7-3630QM for example rq->cpu_capacity_orig is 589 but
> > util_avg scales up to 1024.
> 
> I can't convince myself whether this is the right thing to do. SMT is a
> bit 'special' and it depends on how you model SMT capacity.
> 
> I'm no SMT expert, but the way I understand the current SMT capacity
> model is that capacity_orig represents the capacity of the SMT-thread
> when all its thread-siblings are busy. The true capacity of an
> SMT-thread where all thread-siblings are idle is actually 1024, but we
> don't model this (it would be nightmare to track when the capacity
> should change). The capacity of a core with two or more SMT-threads is
> chosen to be 1024 + smt_gain, where smt_gain is supposed represent the
> additional throughput we gain for the additional SMT-threads. The reason
> why we don't have 1024 per thread is that we would prefer to have only
> one task per core if possible.
> 
> With util_avg scaling to 1024 a core (capacity = 2*589) would be nearly
> 'full' with just one always-running task. If we change util_avg to max
> out at 589, it would take two always-running tasks for the combined
> utilization to match the core capacity. So we may loose some bias
> towards spreading for SMT systems.
> 
> AFAICT, group_is_overloaded() and group_has_capacity() would both be
> affected by this patch.
> 
> Interestingly, Vincent recently proposed to set the SMT-thread capacity
> to 1024 which would affectively make all the current SMT code redundant.
> It would make things a lot simpler, but I'm not sure if we can get away
> with it. It would need discussion at least.
> 
> Opinions?

Thanks for having a look.

The reason I pushed this patch was to address an issue with the
schedutil governor - demand is effectively doubled on SMT systems due to
the above scheme. But this can just be fixed for schedutil by using a
max value there consistent with what __update_load_avg() is using. I'll send
another patch. It looks like there's a good reason for the current PELT
scaling w.r.t. SMT in the scheduler/load balancer.

thanks,
Steve

Re: [PATCH] sched: fix incorrect PELT values on SMT

[Steve Muckle]

On Fri, Aug 19, 2016 at 04:00:57PM +0100, Dietmar Eggemann wrote:
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 61d485421bed..95d34b337152 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -2731,7 +2731,7 @@ __update_load_avg(u64 now, int cpu, struct sched_avg *sa,
> >  	sa->last_update_time = now;
> >  
> >  	scale_freq = arch_scale_freq_capacity(NULL, cpu);
> > -	scale_cpu = arch_scale_cpu_capacity(NULL, cpu);
> > +	scale_cpu = arch_scale_cpu_capacity(cpu_rq(cpu)->sd, cpu);
> 
> Wouldn't you have to subscribe to this rcu pointer rq->sd w/ something
> like 'rcu_dereference(cpu_rq(cpu)->sd)'?
> 
> IMHO, __update_load_avg() is called outside existing RCU read-side
> critical sections as well so there would be a pair of
> rcu_read_lock()/rcu_read_unlock() required in this case.

Thanks Dietmar for the review.

Yeah I didn't consider that this was protected with rcu. It looks like
I'm abandoning this approach anyway though and doing something limited
just to schedutil.

thanks,
Steve

Re: [PATCH] sched: fix incorrect PELT values on SMT

[Wanpeng Li]

2016-08-19 23:30 GMT+08:00 Morten Rasmussen <morten.rasmussen@arm.com>:
> Hi Steve,
>
> On Thu, Aug 18, 2016 at 06:55:41PM -0700, Steve Muckle wrote:
>> PELT scales its util_sum and util_avg values via
>> arch_scale_cpu_capacity(). If that function is passed the CPU's sched
>> domain then it will reduce the scaling capacity if SD_SHARE_CPUCAPACITY
>> is set. PELT does not pass in the sd however. The other caller of
>> arch_scale_cpu_capacity, update_cpu_capacity(), does. This means
>> util_sum and util_avg scale beyond the CPU capacity on SMT.
>>
>> On an Intel i7-3630QM for example rq->cpu_capacity_orig is 589 but
>> util_avg scales up to 1024.
>
> I can't convince myself whether this is the right thing to do. SMT is a
> bit 'special' and it depends on how you model SMT capacity.
>
> I'm no SMT expert, but the way I understand the current SMT capacity
> model is that capacity_orig represents the capacity of the SMT-thread
> when all its thread-siblings are busy. The true capacity of an
> SMT-thread where all thread-siblings are idle is actually 1024, but we
> don't model this (it would be nightmare to track when the capacity
> should change). The capacity of a core with two or more SMT-threads is
> chosen to be 1024 + smt_gain, where smt_gain is supposed represent the
> additional throughput we gain for the additional SMT-threads. The reason
> why we don't have 1024 per thread is that we would prefer to have only
> one task per core if possible.

Agreed, maybe the capacity of an SMP-thread where all thread-siblings
are idle can be 1024 + smt_gain after latest IA technology.
http://www.intel.com/content/www/us/en/architecture-and-technology/turbo-boost/turbo-boost-max-technology.html

Regards,
Wanpeng Li

Re: [PATCH] sched: fix incorrect PELT values on SMT

[Peter Zijlstra]

On Fri, Aug 19, 2016 at 04:30:39PM +0100, Morten Rasmussen wrote:
> I can't convince myself whether this is the right thing to do. SMT is a
> bit 'special' and it depends on how you model SMT capacity.
> 
> I'm no SMT expert, but the way I understand the current SMT capacity
> model is that capacity_orig represents the capacity of the SMT-thread
> when all its thread-siblings are busy.

Correct. Has a weird side effect if you have >2 siblings and unplug some
but not symmetric. Rather uncommon case though.

> The true capacity of an
> SMT-thread where all thread-siblings are idle is actually 1024, but we
> don't model this (it would be nightmare to track when the capacity
> should change). 

Right, so we have some dynamics in the capacity, but doing things like
that (and the power7 asymmetric SMT) requires changing the capacity of
other CPUs, which gets to be real interesting real quick.

The current dynamics are limited to CPU local things, like having RT
tasks eat time.

> The capacity of a core with two or more SMT-threads is
> chosen to be 1024 + smt_gain, where smt_gain is supposed represent the

	(1024 * smt_gain) >> 10

> additional throughput we gain for the additional SMT-threads. The reason
> why we don't have 1024 per thread is that we would prefer to have only
> one task per core if possible.

Not really, it stems from the fact that 1024 used (and still might in
some places) represent 1 (nice-0) task (at 100% utilization).

And if you have SMT you really don't want to stick 2 tasks on if you can
do differently. Simply because 2 threads on a core do not get the same
throughput (in general) as 2 cores do.

Now, these days SD_PREFER_SIBLING might actually be the main force that
gets us 1 task per core if possible. We no longer use the capacity stuff
to compute how many tasks we can run (with exception of
update_numa_stats it seems).

> With util_avg scaling to 1024 a core (capacity = 2*589) would be nearly
> 'full' with just one always-running task. If we change util_avg to max
> out at 589, it would take two always-running tasks for the combined
> utilization to match the core capacity. So we may loose some bias
> towards spreading for SMT systems.

Right, so this is always going to be a bit weird, as util numbers shrink
under load. Therefore they too shrink when you saturate a core with SMT
threads.

> AFAICT, group_is_overloaded() and group_has_capacity() would both be
> affected by this patch.
> 
> Interestingly, Vincent recently proposed to set the SMT-thread capacity
> to 1024 which would affectively make all the current SMT code redundant.
> It would make things a lot simpler, but I'm not sure if we can get away
> with it. It would need discussion at least.
> 
> Opinions?

Time I go stare at SMT again I suppose.. :-)

Re: [PATCH] sched: fix incorrect PELT values on SMT

[Morten Rasmussen]

On Wed, Aug 31, 2016 at 03:07:20PM +0200, Peter Zijlstra wrote:
> On Fri, Aug 19, 2016 at 04:30:39PM +0100, Morten Rasmussen wrote:
> > I can't convince myself whether this is the right thing to do. SMT is a
> > bit 'special' and it depends on how you model SMT capacity.
> > 
> > I'm no SMT expert, but the way I understand the current SMT capacity
> > model is that capacity_orig represents the capacity of the SMT-thread
> > when all its thread-siblings are busy.
> 
> Correct. Has a weird side effect if you have >2 siblings and unplug some
> but not symmetric. Rather uncommon case though.
> 
> > The true capacity of an
> > SMT-thread where all thread-siblings are idle is actually 1024, but we
> > don't model this (it would be nightmare to track when the capacity
> > should change). 
> 
> Right, so we have some dynamics in the capacity, but doing things like
> that (and the power7 asymmetric SMT) requires changing the capacity of
> other CPUs, which gets to be real interesting real quick.
> 
> The current dynamics are limited to CPU local things, like having RT
> tasks eat time.
> 
> > The capacity of a core with two or more SMT-threads is
> > chosen to be 1024 + smt_gain, where smt_gain is supposed represent the
> 
> 	(1024 * smt_gain) >> 10

Looking at the code it seems that we just use smt_gain as the core
capacity, so the SMT capacity is simply sd->smt_gain/sd->span_weight,
where sd->smt_gain is initialized to 1178 by default. But it really
doesn't matter ;-)

> > additional throughput we gain for the additional SMT-threads. The reason
> > why we don't have 1024 per thread is that we would prefer to have only
> > one task per core if possible.
> 
> Not really, it stems from the fact that 1024 used (and still might in
> some places) represent 1 (nice-0) task (at 100% utilization).
> 
> And if you have SMT you really don't want to stick 2 tasks on if you can
> do differently. Simply because 2 threads on a core do not get the same
> throughput (in general) as 2 cores do.

Agreed, that is what I failed to communicate above.

> Now, these days SD_PREFER_SIBLING might actually be the main force that
> gets us 1 task per core if possible. We no longer use the capacity stuff
> to compute how many tasks we can run (with exception of
> update_numa_stats it seems).

Okay. I think the load_above_capacity stuff still does that and we tried
to get rid of that a while back. If we can rely on SD_PREFER_SIBLING
alone, it would certainly make things simpler.

> > With util_avg scaling to 1024 a core (capacity = 2*589) would be nearly
> > 'full' with just one always-running task. If we change util_avg to max
> > out at 589, it would take two always-running tasks for the combined
> > utilization to match the core capacity. So we may loose some bias
> > towards spreading for SMT systems.
> 
> Right, so this is always going to be a bit weird, as util numbers shrink
> under load. Therefore they too shrink when you saturate a core with SMT
> threads.

Shouldn't utilization increase, not shrink, if you saturate more SMT
threads? The effective throughput of each SMT thread should reduce when
more threads are saturated so the utilization should go up since
utilization is time-based?

> > AFAICT, group_is_overloaded() and group_has_capacity() would both be
> > affected by this patch.
> > 
> > Interestingly, Vincent recently proposed to set the SMT-thread capacity
> > to 1024 which would affectively make all the current SMT code redundant.
> > It would make things a lot simpler, but I'm not sure if we can get away
> > with it. It would need discussion at least.
> > 
> > Opinions?
> 
> Time I go stare at SMT again I suppose.. :-)

I'm afraid so.