On Thu, Jan 15, 2015 at 10:09:20AM +0000, Vincent Guittot wrote:
> This patchset consolidates several changes in the capacity and the usage
> tracking of the CPU. It provides a frequency invariant metric of the usage of
> CPUs and generally improves the accuracy of load/usage tracking in the
> scheduler. The frequency invariant metric is the foundation required for the
> consolidation of cpufreq and implementation of a fully invariant load tracking.
> These are currently WIP and require several changes to the load balancer
> (including how it will use and interprets load and capacity metrics) and
> extensive validation. The frequency invariance is done with
> arch_scale_freq_capacity and this patchset doesn't provide the backends of
> the function which are architecture dependent.
>
> As discussed at LPC14, Morten and I have consolidated our changes into a single
> patchset to make it easier to review and merge.
I'm happy with patch 1, 3, 5, 6, and 7. Add my acked-by if you like :)
The last few needs buy-in from somebody running SMT systems I think.
Also, it still not clear why patch 10 uses relative capacity reduction
instead of absolute capacity available to CFS tasks.
Morten
On Thu, Feb 19, 2015 at 12:49:40PM +0000, Morten Rasmussen wrote:
> Also, it still not clear why patch 10 uses relative capacity reduction
> instead of absolute capacity available to CFS tasks.
As present in your asymmetric big and small systems? Yes it would be
unfortunate to migrate a task to an idle small core when the big core is
still faster, even if reduced by rt/irq work.
On Fri, Feb 20, 2015 at 11:34:47AM +0000, Peter Zijlstra wrote:
> On Thu, Feb 19, 2015 at 12:49:40PM +0000, Morten Rasmussen wrote:
>
> > Also, it still not clear why patch 10 uses relative capacity reduction
> > instead of absolute capacity available to CFS tasks.
>
> As present in your asymmetric big and small systems? Yes it would be
> unfortunate to migrate a task to an idle small core when the big core is
> still faster, even if reduced by rt/irq work.
Yes, exactly. I don't think it would cause any harm for symmetric cases
to use absolute capacity instead. Am I missing something?
On 20 February 2015 at 12:52, Morten Rasmussen <[email protected]> wrote:
> On Fri, Feb 20, 2015 at 11:34:47AM +0000, Peter Zijlstra wrote:
>> On Thu, Feb 19, 2015 at 12:49:40PM +0000, Morten Rasmussen wrote:
>>
>> > Also, it still not clear why patch 10 uses relative capacity reduction
>> > instead of absolute capacity available to CFS tasks.
>>
>> As present in your asymmetric big and small systems? Yes it would be
>> unfortunate to migrate a task to an idle small core when the big core is
>> still faster, even if reduced by rt/irq work.
>
> Yes, exactly. I don't think it would cause any harm for symmetric cases
> to use absolute capacity instead. Am I missing something?
If absolute capacity is used, we will trig an active load balance from
little to big core each time a little has got 1 task and a big core is
idle whereas we only want to trig an active migration is the src_cpu's
capacity that is available for the cfs task is significantly reduced
by rt tasks.
I can mix absolute and relative tests by 1st testing that the capacity
of the src is reduced and then ensure that the dst_cpu has more
absolute capacity than src_cpu
On Fri, Feb 20, 2015 at 02:13:21PM +0000, Vincent Guittot wrote:
> On 20 February 2015 at 12:52, Morten Rasmussen <[email protected]> wrote:
> > On Fri, Feb 20, 2015 at 11:34:47AM +0000, Peter Zijlstra wrote:
> >> On Thu, Feb 19, 2015 at 12:49:40PM +0000, Morten Rasmussen wrote:
> >>
> >> > Also, it still not clear why patch 10 uses relative capacity reduction
> >> > instead of absolute capacity available to CFS tasks.
> >>
> >> As present in your asymmetric big and small systems? Yes it would be
> >> unfortunate to migrate a task to an idle small core when the big core is
> >> still faster, even if reduced by rt/irq work.
> >
> > Yes, exactly. I don't think it would cause any harm for symmetric cases
> > to use absolute capacity instead. Am I missing something?
>
> If absolute capacity is used, we will trig an active load balance from
> little to big core each time a little has got 1 task and a big core is
> idle whereas we only want to trig an active migration is the src_cpu's
> capacity that is available for the cfs task is significantly reduced
> by rt tasks.
>
> I can mix absolute and relative tests by 1st testing that the capacity
> of the src is reduced and then ensure that the dst_cpu has more
> absolute capacity than src_cpu
If we use absolute capacity and check if the source cpu is fully
utilized, wouldn't that work? We want to migrate the task if it is
currently being restricted by the available capacity (due to rt/irq
work, being a little cpu, or both) and if there is a destination cpu
with more absolute capacity available. No?
On 20 February 2015 at 15:35, Morten Rasmussen <[email protected]> wrote:
> On Fri, Feb 20, 2015 at 02:13:21PM +0000, Vincent Guittot wrote:
>> On 20 February 2015 at 12:52, Morten Rasmussen <[email protected]> wrote:
>> > On Fri, Feb 20, 2015 at 11:34:47AM +0000, Peter Zijlstra wrote:
>> >> On Thu, Feb 19, 2015 at 12:49:40PM +0000, Morten Rasmussen wrote:
>> >>
>> >> > Also, it still not clear why patch 10 uses relative capacity reduction
>> >> > instead of absolute capacity available to CFS tasks.
>> >>
>> >> As present in your asymmetric big and small systems? Yes it would be
>> >> unfortunate to migrate a task to an idle small core when the big core is
>> >> still faster, even if reduced by rt/irq work.
>> >
>> > Yes, exactly. I don't think it would cause any harm for symmetric cases
>> > to use absolute capacity instead. Am I missing something?
>>
>> If absolute capacity is used, we will trig an active load balance from
>> little to big core each time a little has got 1 task and a big core is
>> idle whereas we only want to trig an active migration is the src_cpu's
>> capacity that is available for the cfs task is significantly reduced
>> by rt tasks.
>>
>> I can mix absolute and relative tests by 1st testing that the capacity
>> of the src is reduced and then ensure that the dst_cpu has more
>> absolute capacity than src_cpu
>
> If we use absolute capacity and check if the source cpu is fully
> utilized, wouldn't that work? We want to migrate the task if it is
we want to trig the migration before the cpu is fully utilized by
rt/irq (which almost never occurs)
> currently being restricted by the available capacity (due to rt/irq
> work, being a little cpu, or both) and if there is a destination cpu
> with more absolute capacity available. No?
yes, so the relative capacity (cpu_capacity vs cpu_capacity_orig)
enables us to know if the cpu is significantly used by irq/rt so it's
worth to do an active load balance of the task. Then the absolute
comparison of cpu_capacity of src_cpu vs cpu_capacity of dst_cpu
checks that the dst_cpu is a better choice
something like :
if ((check_cpu_capacity(src_rq, sd)) &&
(capacity_of(src_cpu)*sd->imbalce_pct < capacity_of(dst_cpu)*100))
return 1;
On Fri, Feb 20, 2015 at 02:54:09PM +0000, Vincent Guittot wrote:
> On 20 February 2015 at 15:35, Morten Rasmussen <[email protected]> wrote:
> > On Fri, Feb 20, 2015 at 02:13:21PM +0000, Vincent Guittot wrote:
> >> On 20 February 2015 at 12:52, Morten Rasmussen <[email protected]> wrote:
> >> > On Fri, Feb 20, 2015 at 11:34:47AM +0000, Peter Zijlstra wrote:
> >> >> On Thu, Feb 19, 2015 at 12:49:40PM +0000, Morten Rasmussen wrote:
> >> >>
> >> >> > Also, it still not clear why patch 10 uses relative capacity reduction
> >> >> > instead of absolute capacity available to CFS tasks.
> >> >>
> >> >> As present in your asymmetric big and small systems? Yes it would be
> >> >> unfortunate to migrate a task to an idle small core when the big core is
> >> >> still faster, even if reduced by rt/irq work.
> >> >
> >> > Yes, exactly. I don't think it would cause any harm for symmetric cases
> >> > to use absolute capacity instead. Am I missing something?
> >>
> >> If absolute capacity is used, we will trig an active load balance from
> >> little to big core each time a little has got 1 task and a big core is
> >> idle whereas we only want to trig an active migration is the src_cpu's
> >> capacity that is available for the cfs task is significantly reduced
> >> by rt tasks.
> >>
> >> I can mix absolute and relative tests by 1st testing that the capacity
> >> of the src is reduced and then ensure that the dst_cpu has more
> >> absolute capacity than src_cpu
> >
> > If we use absolute capacity and check if the source cpu is fully
> > utilized, wouldn't that work? We want to migrate the task if it is
>
> we want to trig the migration before the cpu is fully utilized by
> rt/irq (which almost never occurs)
I meant fully utilized by rt/irq and cfs tasks, sorry. Essentially,
get_cpu_usage() ~= capacity_of(). If get_cpu_usage() is signficantly
smaller than capacity_of() which is may be reduced by rt/irq
utilization, there are still spare cycles and it is not strictly
required to migrate tasks away using active LB. But, tasks would be
moved away if the tasks are being allowed less cpu time due to rt/irq
(get_cpu_usage() >= capacity_of()). Wouldn't that work? Or, do you want
to migrate tasks regardless of whether there are still spare cycles
available on the cpu doing rt/irq work?
The advantage of comparing get_cpu_usage() with capacity_of() is that it
would work for migrating cpu-intensive tasks away from little cpu on
big.LITTLE as well. Then we don't need another almost identical check
for that purpose :)
>
> > currently being restricted by the available capacity (due to rt/irq
> > work, being a little cpu, or both) and if there is a destination cpu
> > with more absolute capacity available. No?
>
> yes, so the relative capacity (cpu_capacity vs cpu_capacity_orig)
> enables us to know if the cpu is significantly used by irq/rt so it's
> worth to do an active load balance of the task. Then the absolute
> comparison of cpu_capacity of src_cpu vs cpu_capacity of dst_cpu
> checks that the dst_cpu is a better choice
>
> something like :
> if ((check_cpu_capacity(src_rq, sd)) &&
> (capacity_of(src_cpu)*sd->imbalce_pct < capacity_of(dst_cpu)*100))
> return 1;
It should solve the big.LITTLE issue. Though I would prefer
get_cpu_usage() ~= capacity_of() approach as it could even improve
performance on big.LITTLE.
On 23 February 2015 at 16:45, Morten Rasmussen <[email protected]> wrote:
> On Fri, Feb 20, 2015 at 02:54:09PM +0000, Vincent Guittot wrote:
>> On 20 February 2015 at 15:35, Morten Rasmussen <[email protected]> wrote:
>> > On Fri, Feb 20, 2015 at 02:13:21PM +0000, Vincent Guittot wrote:
>> >> On 20 February 2015 at 12:52, Morten Rasmussen <[email protected]> wrote:
>> >> > On Fri, Feb 20, 2015 at 11:34:47AM +0000, Peter Zijlstra wrote:
>> >> >> On Thu, Feb 19, 2015 at 12:49:40PM +0000, Morten Rasmussen wrote:
>> >> >>
>> >> >> > Also, it still not clear why patch 10 uses relative capacity reduction
>> >> >> > instead of absolute capacity available to CFS tasks.
>> >> >>
>> >> >> As present in your asymmetric big and small systems? Yes it would be
>> >> >> unfortunate to migrate a task to an idle small core when the big core is
>> >> >> still faster, even if reduced by rt/irq work.
>> >> >
>> >> > Yes, exactly. I don't think it would cause any harm for symmetric cases
>> >> > to use absolute capacity instead. Am I missing something?
>> >>
>> >> If absolute capacity is used, we will trig an active load balance from
>> >> little to big core each time a little has got 1 task and a big core is
>> >> idle whereas we only want to trig an active migration is the src_cpu's
>> >> capacity that is available for the cfs task is significantly reduced
>> >> by rt tasks.
>> >>
>> >> I can mix absolute and relative tests by 1st testing that the capacity
>> >> of the src is reduced and then ensure that the dst_cpu has more
>> >> absolute capacity than src_cpu
>> >
>> > If we use absolute capacity and check if the source cpu is fully
>> > utilized, wouldn't that work? We want to migrate the task if it is
>>
>> we want to trig the migration before the cpu is fully utilized by
>> rt/irq (which almost never occurs)
>
> I meant fully utilized by rt/irq and cfs tasks, sorry. Essentially,
> get_cpu_usage() ~= capacity_of(). If get_cpu_usage() is signficantly
> smaller than capacity_of() which is may be reduced by rt/irq
> utilization, there are still spare cycles and it is not strictly
> required to migrate tasks away using active LB. But, tasks would be
> moved away if the tasks are being allowed less cpu time due to rt/irq
> (get_cpu_usage() >= capacity_of()). Wouldn't that work? Or, do you want
> to migrate tasks regardless of whether there are still spare cycles
> available on the cpu doing rt/irq work?
In fact, we can see perf improvement even if the cpu is not fully used
by thread and interrupts because the task becomes significantly
preempted by interruptions.
>
> The advantage of comparing get_cpu_usage() with capacity_of() is that it
> would work for migrating cpu-intensive tasks away from little cpu on
> big.LITTLE as well. Then we don't need another almost identical check
> for that purpose :)
I understand your point but the patch becomes inefficient for part of
the issue that it's trying to originally solve if we compare
get_cpu_usage with capacity_of. So we will probably need to add few
more tests for the issue you point out above
>
>>
>> > currently being restricted by the available capacity (due to rt/irq
>> > work, being a little cpu, or both) and if there is a destination cpu
>> > with more absolute capacity available. No?
>>
>> yes, so the relative capacity (cpu_capacity vs cpu_capacity_orig)
>> enables us to know if the cpu is significantly used by irq/rt so it's
>> worth to do an active load balance of the task. Then the absolute
>> comparison of cpu_capacity of src_cpu vs cpu_capacity of dst_cpu
>> checks that the dst_cpu is a better choice
>>
>> something like :
>> if ((check_cpu_capacity(src_rq, sd)) &&
>> (capacity_of(src_cpu)*sd->imbalce_pct < capacity_of(dst_cpu)*100))
>> return 1;
>
> It should solve the big.LITTLE issue. Though I would prefer
> get_cpu_usage() ~= capacity_of() approach as it could even improve
> performance on big.LITTLE.
ok. IMHO, it's worth having a dedicated patch for this issue
Vincent
On Tue, Feb 24, 2015 at 10:38:29AM +0000, Vincent Guittot wrote:
> On 23 February 2015 at 16:45, Morten Rasmussen <[email protected]> wrote:
> > On Fri, Feb 20, 2015 at 02:54:09PM +0000, Vincent Guittot wrote:
> >> On 20 February 2015 at 15:35, Morten Rasmussen <[email protected]> wrote:
> >> > On Fri, Feb 20, 2015 at 02:13:21PM +0000, Vincent Guittot wrote:
> >> >> On 20 February 2015 at 12:52, Morten Rasmussen <[email protected]> wrote:
> >> >> > On Fri, Feb 20, 2015 at 11:34:47AM +0000, Peter Zijlstra wrote:
> >> >> >> On Thu, Feb 19, 2015 at 12:49:40PM +0000, Morten Rasmussen wrote:
> >> >> >>
> >> >> >> > Also, it still not clear why patch 10 uses relative capacity reduction
> >> >> >> > instead of absolute capacity available to CFS tasks.
> >> >> >>
> >> >> >> As present in your asymmetric big and small systems? Yes it would be
> >> >> >> unfortunate to migrate a task to an idle small core when the big core is
> >> >> >> still faster, even if reduced by rt/irq work.
> >> >> >
> >> >> > Yes, exactly. I don't think it would cause any harm for symmetric cases
> >> >> > to use absolute capacity instead. Am I missing something?
> >> >>
> >> >> If absolute capacity is used, we will trig an active load balance from
> >> >> little to big core each time a little has got 1 task and a big core is
> >> >> idle whereas we only want to trig an active migration is the src_cpu's
> >> >> capacity that is available for the cfs task is significantly reduced
> >> >> by rt tasks.
> >> >>
> >> >> I can mix absolute and relative tests by 1st testing that the capacity
> >> >> of the src is reduced and then ensure that the dst_cpu has more
> >> >> absolute capacity than src_cpu
> >> >
> >> > If we use absolute capacity and check if the source cpu is fully
> >> > utilized, wouldn't that work? We want to migrate the task if it is
> >>
> >> we want to trig the migration before the cpu is fully utilized by
> >> rt/irq (which almost never occurs)
> >
> > I meant fully utilized by rt/irq and cfs tasks, sorry. Essentially,
> > get_cpu_usage() ~= capacity_of(). If get_cpu_usage() is signficantly
> > smaller than capacity_of() which is may be reduced by rt/irq
> > utilization, there are still spare cycles and it is not strictly
> > required to migrate tasks away using active LB. But, tasks would be
> > moved away if the tasks are being allowed less cpu time due to rt/irq
> > (get_cpu_usage() >= capacity_of()). Wouldn't that work? Or, do you want
> > to migrate tasks regardless of whether there are still spare cycles
> > available on the cpu doing rt/irq work?
>
> In fact, we can see perf improvement even if the cpu is not fully used
> by thread and interrupts because the task becomes significantly
> preempted by interruptions.
Unless the tasks are the consumers of those interrupts, then it would
harm performance to migrate them away :) I get your point though. Could
we have a short comment stating the intentions so we don't forget in a
couple of months?
>
> >
> > The advantage of comparing get_cpu_usage() with capacity_of() is that it
> > would work for migrating cpu-intensive tasks away from little cpu on
> > big.LITTLE as well. Then we don't need another almost identical check
> > for that purpose :)
>
> I understand your point but the patch becomes inefficient for part of
> the issue that it's trying to originally solve if we compare
> get_cpu_usage with capacity_of. So we will probably need to add few
> more tests for the issue you point out above
Right. If your goal is to avoid preemptions and not just make sure that
cpus aren't fully utilized then my proposal isn't sufficient. We will
have to add another condition to solve the big.LITTLE capacity thing
later. In fact we already have that somewhere deep down in the pile of
patches I posted some weeks ago.
> >> > currently being restricted by the available capacity (due to rt/irq
> >> > work, being a little cpu, or both) and if there is a destination cpu
> >> > with more absolute capacity available. No?
> >>
> >> yes, so the relative capacity (cpu_capacity vs cpu_capacity_orig)
> >> enables us to know if the cpu is significantly used by irq/rt so it's
> >> worth to do an active load balance of the task. Then the absolute
> >> comparison of cpu_capacity of src_cpu vs cpu_capacity of dst_cpu
> >> checks that the dst_cpu is a better choice
> >>
> >> something like :
> >> if ((check_cpu_capacity(src_rq, sd)) &&
> >> (capacity_of(src_cpu)*sd->imbalce_pct < capacity_of(dst_cpu)*100))
> >> return 1;
> >
> > It should solve the big.LITTLE issue. Though I would prefer
> > get_cpu_usage() ~= capacity_of() approach as it could even improve
> > performance on big.LITTLE.
>
> ok. IMHO, it's worth having a dedicated patch for this issue
Fine by me as long as we get the extra check you proposed above to fix
the big.LITTLE issue.
Morten
On 24 February 2015 at 12:29, Morten Rasmussen <[email protected]> wrote:
> On Tue, Feb 24, 2015 at 10:38:29AM +0000, Vincent Guittot wrote:
>> On 23 February 2015 at 16:45, Morten Rasmussen <[email protected]> wrote:
>> > On Fri, Feb 20, 2015 at 02:54:09PM +0000, Vincent Guittot wrote:
>> >> On 20 February 2015 at 15:35, Morten Rasmussen <[email protected]> wrote:
>> >> > On Fri, Feb 20, 2015 at 02:13:21PM +0000, Vincent Guittot wrote:
>> >> >> On 20 February 2015 at 12:52, Morten Rasmussen <[email protected]> wrote:
>> >> >> > On Fri, Feb 20, 2015 at 11:34:47AM +0000, Peter Zijlstra wrote:
>> >> >> >> On Thu, Feb 19, 2015 at 12:49:40PM +0000, Morten Rasmussen wrote:
>> >> >> >>
>> >> >> >> > Also, it still not clear why patch 10 uses relative capacity reduction
>> >> >> >> > instead of absolute capacity available to CFS tasks.
>> >> >> >>
>> >> >> >> As present in your asymmetric big and small systems? Yes it would be
>> >> >> >> unfortunate to migrate a task to an idle small core when the big core is
>> >> >> >> still faster, even if reduced by rt/irq work.
>> >> >> >
>> >> >> > Yes, exactly. I don't think it would cause any harm for symmetric cases
>> >> >> > to use absolute capacity instead. Am I missing something?
>> >> >>
>> >> >> If absolute capacity is used, we will trig an active load balance from
>> >> >> little to big core each time a little has got 1 task and a big core is
>> >> >> idle whereas we only want to trig an active migration is the src_cpu's
>> >> >> capacity that is available for the cfs task is significantly reduced
>> >> >> by rt tasks.
>> >> >>
>> >> >> I can mix absolute and relative tests by 1st testing that the capacity
>> >> >> of the src is reduced and then ensure that the dst_cpu has more
>> >> >> absolute capacity than src_cpu
>> >> >
>> >> > If we use absolute capacity and check if the source cpu is fully
>> >> > utilized, wouldn't that work? We want to migrate the task if it is
>> >>
>> >> we want to trig the migration before the cpu is fully utilized by
>> >> rt/irq (which almost never occurs)
>> >
>> > I meant fully utilized by rt/irq and cfs tasks, sorry. Essentially,
>> > get_cpu_usage() ~= capacity_of(). If get_cpu_usage() is signficantly
>> > smaller than capacity_of() which is may be reduced by rt/irq
>> > utilization, there are still spare cycles and it is not strictly
>> > required to migrate tasks away using active LB. But, tasks would be
>> > moved away if the tasks are being allowed less cpu time due to rt/irq
>> > (get_cpu_usage() >= capacity_of()). Wouldn't that work? Or, do you want
>> > to migrate tasks regardless of whether there are still spare cycles
>> > available on the cpu doing rt/irq work?
>>
>> In fact, we can see perf improvement even if the cpu is not fully used
>> by thread and interrupts because the task becomes significantly
>> preempted by interruptions.
>
> Unless the tasks are the consumers of those interrupts, then it would
> harm performance to migrate them away :) I get your point though. Could
> we have a short comment stating the intentions so we don't forget in a
> couple of months?
I will add more details in the commit log
>
>>
>> >
>> > The advantage of comparing get_cpu_usage() with capacity_of() is that it
>> > would work for migrating cpu-intensive tasks away from little cpu on
>> > big.LITTLE as well. Then we don't need another almost identical check
>> > for that purpose :)
>>
>> I understand your point but the patch becomes inefficient for part of
>> the issue that it's trying to originally solve if we compare
>> get_cpu_usage with capacity_of. So we will probably need to add few
>> more tests for the issue you point out above
>
> Right. If your goal is to avoid preemptions and not just make sure that
> cpus aren't fully utilized then my proposal isn't sufficient. We will
> have to add another condition to solve the big.LITTLE capacity thing
> later. In fact we already have that somewhere deep down in the pile of
> patches I posted some weeks ago.
>
>> >> > currently being restricted by the available capacity (due to rt/irq
>> >> > work, being a little cpu, or both) and if there is a destination cpu
>> >> > with more absolute capacity available. No?
>> >>
>> >> yes, so the relative capacity (cpu_capacity vs cpu_capacity_orig)
>> >> enables us to know if the cpu is significantly used by irq/rt so it's
>> >> worth to do an active load balance of the task. Then the absolute
>> >> comparison of cpu_capacity of src_cpu vs cpu_capacity of dst_cpu
>> >> checks that the dst_cpu is a better choice
>> >>
>> >> something like :
>> >> if ((check_cpu_capacity(src_rq, sd)) &&
>> >> (capacity_of(src_cpu)*sd->imbalce_pct < capacity_of(dst_cpu)*100))
>> >> return 1;
>> >
>> > It should solve the big.LITTLE issue. Though I would prefer
>> > get_cpu_usage() ~= capacity_of() approach as it could even improve
>> > performance on big.LITTLE.
>>
>> ok. IMHO, it's worth having a dedicated patch for this issue
>
> Fine by me as long as we get the extra check you proposed above to fix
> the big.LITTLE issue.
ok
>
> Morten