Problem statement:
Once the user has disabled turbo frequency by
echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo,
the cfs_rq's util_avg becomes quite small when compared with
CPU capacity.
Step to reproduce:
echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo
./x86_cpuload --count 1 --start 3 --timeout 100 --busy 99
would launch 1 thread and bind it to CPU3, lasting for 100 seconds,
with a CPU utilization of 99%. [1]
top result:
%Cpu3 : 98.4 us, 0.0 sy, 0.0 ni, 1.6 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
check util_avg:
cat /sys/kernel/debug/sched/debug | grep "cfs_rq\[3\]" -A 20 | grep util_avg
.util_avg : 611
So the util_avg/cpu capacity is 611/1024, which is much smaller than
98.4% shown in the top result.
This might impact some logic in the scheduler. For example, group_is_overloaded()
would compare the group_capacity and group_util in the sched group, to
check if this sched group is overloaded or not. With this gap, even
when there is a nearly 100% workload, the sched group will not be regarded
as overloaded. Besides group_is_overloaded(), there are also other victims.
There is a ongoing work that aims to optimize the task wakeup in a LLC domain.
The main idea is to stop searching idle CPUs if the sched domain is overloaded[2].
This proposal also relies on the util_avg/CPU capacity to decide whether the LLC
domain is overloaded.
Analysis:
CPU frequency invariance has caused this difference. In summary,
the util_sum of cfs rq would decay quite fast when the CPU is in
idle, when the CPU frequency invariance is enabled.
The detail is as followed:
As depicted in update_rq_clock_pelt(), when the frequency invariance
is enabled, there would be two clock variables on each rq, clock_task
and clock_pelt:
The clock_pelt scales the time to reflect the effective amount of
computation done during the running delta time but then syncs back to
clock_task when rq is idle.
absolute time | 1| 2| 3| 4| 5| 6| 7| 8| 9|10|11|12|13|14|15|16
@ max frequency ------******---------------******---------------
@ half frequency ------************---------************---------
clock pelt | 1| 2| 3| 4| 7| 8| 9| 10| 11|14|15|16
The fast decay of util_sum during idle is due to:
1. rq->clock_pelt is always behind rq->clock_task
2. rq->last_update is updated to rq->clock_pelt' after invoking ___update_load_sum()
3. Then the CPU becomes idle, the rq->clock_pelt' would be suddenly increased
a lot to rq->clock_task
4. Enters ___update_load_sum() again, the idle period is calculated by
rq->clock_task - rq->last_update, AKA, rq->clock_task - rq->clock_pelt'.
The lower the CPU frequency is, the larger the delta =
rq->clock_task - rq->clock_pelt' will be. Since the idle period will be
used to decay the util_sum only, the util_sum drops significantly during
idle period.
Proposal:
This symptom is not only caused by disabling turbo frequency, but it
would also appear if the user limits the max frequency at runtime. Because
if the frequency is always lower than the max frequency,
CPU frequency invariance would decay the util_sum quite fast during idle.
As some end users would disable turbo after boot up, this patch aims to
present this symptom and deals with turbo scenarios for now. It might
be ideal if CPU frequency invariance is aware of the max CPU frequency
(user specified) at runtime in the future.
[Previous patch seems to be lost on LKML, this is a resend, sorry for any
inconvenience]
Link: https://github.com/yu-chen-surf/x86_cpuload.git #1
Link: https://lore.kernel.org/lkml/[email protected]/ #2
Signed-off-by: Chen Yu <[email protected]>
---
drivers/cpufreq/intel_pstate.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
index 846bb3a78788..2216b24b6f84 100644
--- a/drivers/cpufreq/intel_pstate.c
+++ b/drivers/cpufreq/intel_pstate.c
@@ -1322,6 +1322,7 @@ static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
mutex_unlock(&intel_pstate_limits_lock);
intel_pstate_update_policies();
+ arch_set_max_freq_ratio(global.no_turbo);
mutex_unlock(&intel_pstate_driver_lock);
--
2.25.1
On Fri, Apr 08, 2022 at 07:42:58AM +0800, Chen Yu wrote:
> Problem statement:
> Once the user has disabled turbo frequency by
> echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo,
> the cfs_rq's util_avg becomes quite small when compared with
> CPU capacity.
>
> Step to reproduce:
>
> echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo
>
> ./x86_cpuload --count 1 --start 3 --timeout 100 --busy 99
> would launch 1 thread and bind it to CPU3, lasting for 100 seconds,
> with a CPU utilization of 99%. [1]
>
> top result:
> %Cpu3 : 98.4 us, 0.0 sy, 0.0 ni, 1.6 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
>
> check util_avg:
> cat /sys/kernel/debug/sched/debug | grep "cfs_rq\[3\]" -A 20 | grep util_avg
> .util_avg : 611
>
> So the util_avg/cpu capacity is 611/1024, which is much smaller than
> 98.4% shown in the top result.
>
> This might impact some logic in the scheduler. For example, group_is_overloaded()
> would compare the group_capacity and group_util in the sched group, to
> check if this sched group is overloaded or not. With this gap, even
> when there is a nearly 100% workload, the sched group will not be regarded
> as overloaded. Besides group_is_overloaded(), there are also other victims.
> There is a ongoing work that aims to optimize the task wakeup in a LLC domain.
> The main idea is to stop searching idle CPUs if the sched domain is overloaded[2].
> This proposal also relies on the util_avg/CPU capacity to decide whether the LLC
> domain is overloaded.
>
> Analysis:
> CPU frequency invariance has caused this difference. In summary,
> the util_sum of cfs rq would decay quite fast when the CPU is in
> idle, when the CPU frequency invariance is enabled.
>
> The detail is as followed:
>
> As depicted in update_rq_clock_pelt(), when the frequency invariance
> is enabled, there would be two clock variables on each rq, clock_task
> and clock_pelt:
>
> The clock_pelt scales the time to reflect the effective amount of
> computation done during the running delta time but then syncs back to
> clock_task when rq is idle.
>
> absolute time | 1| 2| 3| 4| 5| 6| 7| 8| 9|10|11|12|13|14|15|16
> @ max frequency ------******---------------******---------------
> @ half frequency ------************---------************---------
> clock pelt | 1| 2| 3| 4| 7| 8| 9| 10| 11|14|15|16
>
> The fast decay of util_sum during idle is due to:
> 1. rq->clock_pelt is always behind rq->clock_task
> 2. rq->last_update is updated to rq->clock_pelt' after invoking ___update_load_sum()
> 3. Then the CPU becomes idle, the rq->clock_pelt' would be suddenly increased
> a lot to rq->clock_task
> 4. Enters ___update_load_sum() again, the idle period is calculated by
> rq->clock_task - rq->last_update, AKA, rq->clock_task - rq->clock_pelt'.
> The lower the CPU frequency is, the larger the delta =
> rq->clock_task - rq->clock_pelt' will be. Since the idle period will be
> used to decay the util_sum only, the util_sum drops significantly during
> idle period.
>
> Proposal:
> This symptom is not only caused by disabling turbo frequency, but it
> would also appear if the user limits the max frequency at runtime. Because
> if the frequency is always lower than the max frequency,
> CPU frequency invariance would decay the util_sum quite fast during idle.
>
> As some end users would disable turbo after boot up, this patch aims to
> present this symptom and deals with turbo scenarios for now. It might
> be ideal if CPU frequency invariance is aware of the max CPU frequency
> (user specified) at runtime in the future.
>
> [Previous patch seems to be lost on LKML, this is a resend, sorry for any
> inconvenience]
>
> Link: https://github.com/yu-chen-surf/x86_cpuload.git #1
> Link: https://lore.kernel.org/lkml/[email protected]/ #2
> Signed-off-by: Chen Yu <[email protected]>
Acked-by: Peter Zijlstra (Intel) <[email protected]>
On Fri, 2022-04-08 at 07:42 +0800, Chen Yu wrote:
> Problem statement:
> Once the user has disabled turbo frequency by
> echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo,
> the cfs_rq's util_avg becomes quite small when compared with
> CPU capacity.
>
> Step to reproduce:
>
> echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo
>
> ./x86_cpuload --count 1 --start 3 --timeout 100 --busy 99
> would launch 1 thread and bind it to CPU3, lasting for 100 seconds,
> with a CPU utilization of 99%. [1]
>
> top result:
> %Cpu3 : 98.4 us, 0.0 sy, 0.0 ni, 1.6 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
>
> check util_avg:
> cat /sys/kernel/debug/sched/debug | grep "cfs_rq\[3\]" -A 20 | grep util_avg
> .util_avg : 611
>
> So the util_avg/cpu capacity is 611/1024, which is much smaller than
> 98.4% shown in the top result.
>
> This might impact some logic in the scheduler. For example, group_is_overloaded()
> would compare the group_capacity and group_util in the sched group, to
> check if this sched group is overloaded or not. With this gap, even
> when there is a nearly 100% workload, the sched group will not be regarded
> as overloaded. Besides group_is_overloaded(), there are also other victims.
> There is a ongoing work that aims to optimize the task wakeup in a LLC domain.
> The main idea is to stop searching idle CPUs if the sched domain is overloaded[2].
> This proposal also relies on the util_avg/CPU capacity to decide whether the LLC
> domain is overloaded.
>
> Analysis:
> CPU frequency invariance has caused this difference. In summary,
> the util_sum of cfs rq would decay quite fast when the CPU is in
> idle, when the CPU frequency invariance is enabled.
>
> The detail is as followed:
>
> As depicted in update_rq_clock_pelt(), when the frequency invariance
> is enabled, there would be two clock variables on each rq, clock_task
> and clock_pelt:
>
> The clock_pelt scales the time to reflect the effective amount of
> computation done during the running delta time but then syncs back to
> clock_task when rq is idle.
>
> absolute time | 1| 2| 3| 4| 5| 6| 7| 8| 9|10|11|12|13|14|15|16
> @ max frequency ------******---------------******---------------
> @ half frequency ------************---------************---------
> clock pelt | 1| 2| 3| 4| 7| 8| 9| 10| 11|14|15|16
>
> The fast decay of util_sum during idle is due to:
> 1. rq->clock_pelt is always behind rq->clock_task
> 2. rq->last_update is updated to rq->clock_pelt' after invoking ___update_load_sum()
> 3. Then the CPU becomes idle, the rq->clock_pelt' would be suddenly increased
> a lot to rq->clock_task
> 4. Enters ___update_load_sum() again, the idle period is calculated by
> rq->clock_task - rq->last_update, AKA, rq->clock_task - rq->clock_pelt'.
> The lower the CPU frequency is, the larger the delta =
> rq->clock_task - rq->clock_pelt' will be. Since the idle period will be
> used to decay the util_sum only, the util_sum drops significantly during
> idle period.
>
> Proposal:
> This symptom is not only caused by disabling turbo frequency, but it
> would also appear if the user limits the max frequency at runtime. Because
> if the frequency is always lower than the max frequency,
> CPU frequency invariance would decay the util_sum quite fast during idle.
>
> As some end users would disable turbo after boot up, this patch aims to
> present this symptom and deals with turbo scenarios for now. It might
> be ideal if CPU frequency invariance is aware of the max CPU frequency
> (user specified) at runtime in the future.
>
> [Previous patch seems to be lost on LKML, this is a resend, sorry for any
> inconvenience]
>
> Link: https://github.com/yu-chen-surf/x86_cpuload.git #1
> Link: https://lore.kernel.org/lkml/[email protected]/ #2
> Signed-off-by: Chen Yu <[email protected]>
Reviewed-by: Giovanni Gherdovich <[email protected]>
You're right, when turbo is disabled, the frequency invariance code needs to
know about it; it calculates freq_curr/freq_max and thinks that freq_max is
some turbo level. For example commit 918229cdd5ab ("x86/intel_pstate: Handle
runtime turbo disablement/enablement in frequency invariance") takes care of
this when global.turbo_disabled changes, but before your patch nothing checks
if the user disabled turbo from sysfs. Thanks for the fix!
Giovanni
> ---
> drivers/cpufreq/intel_pstate.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
> index 846bb3a78788..2216b24b6f84 100644
> --- a/drivers/cpufreq/intel_pstate.c
> +++ b/drivers/cpufreq/intel_pstate.c
> @@ -1322,6 +1322,7 @@ static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
> mutex_unlock(&intel_pstate_limits_lock);
>
> intel_pstate_update_policies();
> + arch_set_max_freq_ratio(global.no_turbo);
>
> mutex_unlock(&intel_pstate_driver_lock);
>
On Fri, Apr 8, 2022 at 4:22 PM Giovanni Gherdovich <[email protected]> wrote:
>
> On Fri, 2022-04-08 at 07:42 +0800, Chen Yu wrote:
> > Problem statement:
> > Once the user has disabled turbo frequency by
> > echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo,
> > the cfs_rq's util_avg becomes quite small when compared with
> > CPU capacity.
> >
> > Step to reproduce:
> >
> > echo 1 > /sys/devices/system/cpu/intel_pstate/no_turbo
> >
> > ./x86_cpuload --count 1 --start 3 --timeout 100 --busy 99
> > would launch 1 thread and bind it to CPU3, lasting for 100 seconds,
> > with a CPU utilization of 99%. [1]
> >
> > top result:
> > %Cpu3 : 98.4 us, 0.0 sy, 0.0 ni, 1.6 id, 0.0 wa, 0.0 hi, 0.0 si, 0.0 st
> >
> > check util_avg:
> > cat /sys/kernel/debug/sched/debug | grep "cfs_rq\[3\]" -A 20 | grep util_avg
> > .util_avg : 611
> >
> > So the util_avg/cpu capacity is 611/1024, which is much smaller than
> > 98.4% shown in the top result.
> >
> > This might impact some logic in the scheduler. For example, group_is_overloaded()
> > would compare the group_capacity and group_util in the sched group, to
> > check if this sched group is overloaded or not. With this gap, even
> > when there is a nearly 100% workload, the sched group will not be regarded
> > as overloaded. Besides group_is_overloaded(), there are also other victims.
> > There is a ongoing work that aims to optimize the task wakeup in a LLC domain.
> > The main idea is to stop searching idle CPUs if the sched domain is overloaded[2].
> > This proposal also relies on the util_avg/CPU capacity to decide whether the LLC
> > domain is overloaded.
> >
> > Analysis:
> > CPU frequency invariance has caused this difference. In summary,
> > the util_sum of cfs rq would decay quite fast when the CPU is in
> > idle, when the CPU frequency invariance is enabled.
> >
> > The detail is as followed:
> >
> > As depicted in update_rq_clock_pelt(), when the frequency invariance
> > is enabled, there would be two clock variables on each rq, clock_task
> > and clock_pelt:
> >
> > The clock_pelt scales the time to reflect the effective amount of
> > computation done during the running delta time but then syncs back to
> > clock_task when rq is idle.
> >
> > absolute time | 1| 2| 3| 4| 5| 6| 7| 8| 9|10|11|12|13|14|15|16
> > @ max frequency ------******---------------******---------------
> > @ half frequency ------************---------************---------
> > clock pelt | 1| 2| 3| 4| 7| 8| 9| 10| 11|14|15|16
> >
> > The fast decay of util_sum during idle is due to:
> > 1. rq->clock_pelt is always behind rq->clock_task
> > 2. rq->last_update is updated to rq->clock_pelt' after invoking ___update_load_sum()
> > 3. Then the CPU becomes idle, the rq->clock_pelt' would be suddenly increased
> > a lot to rq->clock_task
> > 4. Enters ___update_load_sum() again, the idle period is calculated by
> > rq->clock_task - rq->last_update, AKA, rq->clock_task - rq->clock_pelt'.
> > The lower the CPU frequency is, the larger the delta =
> > rq->clock_task - rq->clock_pelt' will be. Since the idle period will be
> > used to decay the util_sum only, the util_sum drops significantly during
> > idle period.
> >
> > Proposal:
> > This symptom is not only caused by disabling turbo frequency, but it
> > would also appear if the user limits the max frequency at runtime. Because
> > if the frequency is always lower than the max frequency,
> > CPU frequency invariance would decay the util_sum quite fast during idle.
> >
> > As some end users would disable turbo after boot up, this patch aims to
> > present this symptom and deals with turbo scenarios for now. It might
> > be ideal if CPU frequency invariance is aware of the max CPU frequency
> > (user specified) at runtime in the future.
> >
> > [Previous patch seems to be lost on LKML, this is a resend, sorry for any
> > inconvenience]
> >
> > Link: https://github.com/yu-chen-surf/x86_cpuload.git #1
> > Link: https://lore.kernel.org/lkml/[email protected]/ #2
> > Signed-off-by: Chen Yu <[email protected]>
>
> Reviewed-by: Giovanni Gherdovich <[email protected]>
>
> You're right, when turbo is disabled, the frequency invariance code needs to
> know about it; it calculates freq_curr/freq_max and thinks that freq_max is
> some turbo level. For example commit 918229cdd5ab ("x86/intel_pstate: Handle
> runtime turbo disablement/enablement in frequency invariance") takes care of
> this when global.turbo_disabled changes, but before your patch nothing checks
> if the user disabled turbo from sysfs. Thanks for the fix!
>
> Giovanni
Applied as 5.19 material, thanks!