| Message ID | 20210604080954.13915-2-lukasz.luba@arm.com |
|---|---|
| State | New |
| Headers | show |
| Series | Add allowed CPU capacity knowledge to EAS | expand |
On Fri, 4 Jun 2021 at 10:10, Lukasz Luba <lukasz.luba@arm.com> wrote: > > Energy Aware Scheduling (EAS) needs to be able to predict the frequency > requests made by the SchedUtil governor to properly estimate energy used > in the future. It has to take into account CPUs utilization and forecast > Performance Domain (PD) frequency. There is a corner case when the max > allowed frequency might be reduced due to thermal. SchedUtil is aware of > that reduced frequency, so it should be taken into account also in EAS > estimations. > > SchedUtil, as a CPUFreq governor, knows the maximum allowed frequency of > a CPU, thanks to cpufreq_driver_resolve_freq() and internal clamping > to 'policy::max'. SchedUtil is responsible to respect that upper limit > while setting the frequency through CPUFreq drivers. This effective > frequency is stored internally in 'sugov_policy::next_freq' and EAS has > to predict that value. > > In the existing code the raw value of arch_scale_cpu_capacity() is used > for clamping the returned CPU utilization from effective_cpu_util(). > This patch fixes issue with too big single CPU utilization, by introducing > clamping to the allowed CPU capacity. The allowed CPU capacity is a CPU > capacity reduced by thermal pressure signal. We rely on this load avg > geometric series in similar way as other mechanisms in the scheduler. > > Thanks to knowledge about allowed CPU capacity, we don't get too big value > for a single CPU utilization, which is then added to the util sum. The > util sum is used as a source of information for estimating whole PD energy. > To avoid wrong energy estimation in EAS (due to capped frequency), make > sure that the calculation of util sum is aware of allowed CPU capacity. > > Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> > --- > kernel/sched/fair.c | 17 ++++++++++++++--- > 1 file changed, 14 insertions(+), 3 deletions(-) > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c > index 161b92aa1c79..1aeddecabc20 100644 > --- a/kernel/sched/fair.c > +++ b/kernel/sched/fair.c > @@ -6527,6 +6527,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > struct cpumask *pd_mask = perf_domain_span(pd); > unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); > unsigned long max_util = 0, sum_util = 0; > + unsigned long _cpu_cap = cpu_cap; > int cpu; > > /* > @@ -6558,14 +6559,24 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > cpu_util_next(cpu, p, -1) + task_util_est(p); > } > > + /* > + * Take the thermal pressure from non-idle CPUs. They have > + * most up-to-date information. For idle CPUs thermal pressure > + * signal is not updated so often. What do you mean by "not updated so often" ? Do you have a value ? Thermal pressure is updated at the same rate as other PELT values of an idle CPU. Why is it a problem there ? > + */ > + if (!idle_cpu(cpu)) > + _cpu_cap = cpu_cap - thermal_load_avg(cpu_rq(cpu)); > + > /* > * Busy time computation: utilization clamping is not > * required since the ratio (sum_util / cpu_capacity) > * is already enough to scale the EM reported power > * consumption at the (eventually clamped) cpu_capacity. > */ > - sum_util += effective_cpu_util(cpu, util_running, cpu_cap, > - ENERGY_UTIL, NULL); > + cpu_util = effective_cpu_util(cpu, util_running, cpu_cap, > + ENERGY_UTIL, NULL); > + > + sum_util += min(cpu_util, _cpu_cap); > > /* > * Performance domain frequency: utilization clamping > @@ -6576,7 +6587,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > */ > cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap, > FREQUENCY_UTIL, tsk); > - max_util = max(max_util, cpu_util); > + max_util = max(max_util, min(cpu_util, _cpu_cap)); > } > > return em_cpu_energy(pd->em_pd, max_util, sum_util); > -- > 2.17.1 >
On 04/06/2021 10:09, Lukasz Luba wrote: > Energy Aware Scheduling (EAS) needs to be able to predict the frequency > requests made by the SchedUtil governor to properly estimate energy used > in the future. It has to take into account CPUs utilization and forecast > Performance Domain (PD) frequency. There is a corner case when the max > allowed frequency might be reduced due to thermal. SchedUtil is aware of > that reduced frequency, so it should be taken into account also in EAS > estimations. > > SchedUtil, as a CPUFreq governor, knows the maximum allowed frequency of > a CPU, thanks to cpufreq_driver_resolve_freq() and internal clamping > to 'policy::max'. SchedUtil is responsible to respect that upper limit > while setting the frequency through CPUFreq drivers. This effective > frequency is stored internally in 'sugov_policy::next_freq' and EAS has > to predict that value. > > In the existing code the raw value of arch_scale_cpu_capacity() is used > for clamping the returned CPU utilization from effective_cpu_util(). > This patch fixes issue with too big single CPU utilization, by introducing > clamping to the allowed CPU capacity. The allowed CPU capacity is a CPU > capacity reduced by thermal pressure signal. We rely on this load avg > geometric series in similar way as other mechanisms in the scheduler. > > Thanks to knowledge about allowed CPU capacity, we don't get too big value > for a single CPU utilization, which is then added to the util sum. The > util sum is used as a source of information for estimating whole PD energy. > To avoid wrong energy estimation in EAS (due to capped frequency), make > sure that the calculation of util sum is aware of allowed CPU capacity. > So essentially what you want to do is: Make EAS aware of the frequency clamping schedutil can be faced with: get_next_freq() -> cpufreq_driver_resolve_freq() -> clamp_val(target_freq, policy->min, policy->max) (1) by subtracting CPU's Thermal Pressure (ThPr) signal from the original CPU capacity `arch_scale_cpu_capacity()` (2). --- Isn't there a conceptional flaw in this design? Let's say we have a big.Little system with two cpufreq cooling devices and a thermal zone (something like Hikey 960). To create a ThPr scenario we have to run stuff on the CPUs (e.g. hackbench (3)). Eventually cpufreq_set_cur_state() [drivers/thermal/cpufreq_cooling.c] will set thermal_pressure to `(2) - (2)*freq/policy->cpuinfo.max_freq` and PELT will provide the ThPr signal via thermal_load_avg(). But to create this scenario, the system will become overutilized (system-wide data, if one CPU is overutilized, the whole system is) so EAS is disabled (i.e. find_energy_efficient_cpu() and compute_emergy() are not executed). I can see that there are episodes in which EAS is running and thermal_load_avg() != 0 but those have to be when (3) has stopped and you see the ThPr signal just decaying (no accruing of new ThPr). The cpufreq cooling device can still issue cpufreq_set_cur_state() but only with decreasing states. --- IMHO, a precise description of how you envision the system setup, incorporating all participating subsystems, would be helpful here. > Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> > --- > kernel/sched/fair.c | 17 ++++++++++++++--- > 1 file changed, 14 insertions(+), 3 deletions(-) > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c > index 161b92aa1c79..1aeddecabc20 100644 > --- a/kernel/sched/fair.c > +++ b/kernel/sched/fair.c > @@ -6527,6 +6527,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > struct cpumask *pd_mask = perf_domain_span(pd); > unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); > unsigned long max_util = 0, sum_util = 0; > + unsigned long _cpu_cap = cpu_cap; > int cpu; > > /* > @@ -6558,14 +6559,24 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > cpu_util_next(cpu, p, -1) + task_util_est(p); > } > > + /* > + * Take the thermal pressure from non-idle CPUs. They have > + * most up-to-date information. For idle CPUs thermal pressure > + * signal is not updated so often. > + */ > + if (!idle_cpu(cpu)) > + _cpu_cap = cpu_cap - thermal_load_avg(cpu_rq(cpu)); > + This one is probably the result of the fact that cpufreq cooling device sets the ThPr for all CPUs of the policy (Frequency Domain (FD) or Performance Domain (PD)) but PELT updates are happening per-CPU. And only !idle CPUs get the update in scheduler_tick(). Looks like thermal_pressure [per_cpu(thermal_pressure, cpu), drivers/base/arch_topology.c] set by cpufreq_set_cur_state() is always in sync with policy->max/cpuinfo_max_freq). So for your use case this instantaneous `signal` is better than the PELT one. It's precise (no decaying when frequency clamping is already gone) and you avoid the per-cpu update issue. > /* > * Busy time computation: utilization clamping is not > * required since the ratio (sum_util / cpu_capacity) > * is already enough to scale the EM reported power > * consumption at the (eventually clamped) cpu_capacity. > */ > - sum_util += effective_cpu_util(cpu, util_running, cpu_cap, > - ENERGY_UTIL, NULL); > + cpu_util = effective_cpu_util(cpu, util_running, cpu_cap, > + ENERGY_UTIL, NULL); > + > + sum_util += min(cpu_util, _cpu_cap); > > /* > * Performance domain frequency: utilization clamping > @@ -6576,7 +6587,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > */ > cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap, > FREQUENCY_UTIL, tsk); > - max_util = max(max_util, cpu_util); > + max_util = max(max_util, min(cpu_util, _cpu_cap)); > } > > return em_cpu_energy(pd->em_pd, max_util, sum_util); >
On 6/10/21 8:59 AM, Vincent Guittot wrote: > On Fri, 4 Jun 2021 at 10:10, Lukasz Luba <lukasz.luba@arm.com> wrote: >> >> Energy Aware Scheduling (EAS) needs to be able to predict the frequency >> requests made by the SchedUtil governor to properly estimate energy used >> in the future. It has to take into account CPUs utilization and forecast >> Performance Domain (PD) frequency. There is a corner case when the max >> allowed frequency might be reduced due to thermal. SchedUtil is aware of >> that reduced frequency, so it should be taken into account also in EAS >> estimations. >> >> SchedUtil, as a CPUFreq governor, knows the maximum allowed frequency of >> a CPU, thanks to cpufreq_driver_resolve_freq() and internal clamping >> to 'policy::max'. SchedUtil is responsible to respect that upper limit >> while setting the frequency through CPUFreq drivers. This effective >> frequency is stored internally in 'sugov_policy::next_freq' and EAS has >> to predict that value. >> >> In the existing code the raw value of arch_scale_cpu_capacity() is used >> for clamping the returned CPU utilization from effective_cpu_util(). >> This patch fixes issue with too big single CPU utilization, by introducing >> clamping to the allowed CPU capacity. The allowed CPU capacity is a CPU >> capacity reduced by thermal pressure signal. We rely on this load avg >> geometric series in similar way as other mechanisms in the scheduler. >> >> Thanks to knowledge about allowed CPU capacity, we don't get too big value >> for a single CPU utilization, which is then added to the util sum. The >> util sum is used as a source of information for estimating whole PD energy. >> To avoid wrong energy estimation in EAS (due to capped frequency), make >> sure that the calculation of util sum is aware of allowed CPU capacity. >> >> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> >> --- >> kernel/sched/fair.c | 17 ++++++++++++++--- >> 1 file changed, 14 insertions(+), 3 deletions(-) >> >> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c >> index 161b92aa1c79..1aeddecabc20 100644 >> --- a/kernel/sched/fair.c >> +++ b/kernel/sched/fair.c >> @@ -6527,6 +6527,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) >> struct cpumask *pd_mask = perf_domain_span(pd); >> unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); >> unsigned long max_util = 0, sum_util = 0; >> + unsigned long _cpu_cap = cpu_cap; >> int cpu; >> >> /* >> @@ -6558,14 +6559,24 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) >> cpu_util_next(cpu, p, -1) + task_util_est(p); >> } >> >> + /* >> + * Take the thermal pressure from non-idle CPUs. They have >> + * most up-to-date information. For idle CPUs thermal pressure >> + * signal is not updated so often. > > What do you mean by "not updated so often" ? Do you have a value ? > > Thermal pressure is updated at the same rate as other PELT values of > an idle CPU. Why is it a problem there ? > For idle CPU the value is updated 'remotely' by some other CPU running nohz_idle_balance(). That goes into update_blocked_averages() if the flags and checks are OK inside update_nohz_stats(). Sometimes this is not called because other_have_blocked() returned false. It can happen for a long idle CPU, which all signals in that function has 0 [1]. This will cause that we don't check what is a new value stored by thermal cpufreq_cooling for the thermal pressure [2]. We should feed that value into the 'signal' machinery inside the __update_blocked_others() [3]. Unfortunately, in a corner case there's a flag (rq->has_blocked_load) which blocks the check of a raw thermal value and prevents feeding it into thermal pressure signal (since it's a long idle CPU, there is no load) [4]. It has implication on this patch, because I cannot e.g. take first CPU from the PD mask and blindly check it's thermal pressure, because it can be idle for a long time. I don't want to have two loop, first just for taking the latest thermal pressure for the PD. Thus, I want to re-use the existing loop to take the latest information from non-idle CPU and pass use. Regards, Lukasz [1] https://elixir.bootlin.com/linux/latest/source/kernel/sched/fair.c#L7909 [2] https://elixir.bootlin.com/linux/latest/source/drivers/thermal/cpufreq_cooling.c#L494 [3] https://elixir.bootlin.com/linux/latest/source/kernel/sched/fair.c#L7958 [4] https://elixir.bootlin.com/linux/latest/source/kernel/sched/fair.c#L8433
On 6/10/21 9:42 AM, Dietmar Eggemann wrote: [snip] > > So essentially what you want to do is: > > Make EAS aware of the frequency clamping schedutil can be faced with: > > get_next_freq() -> cpufreq_driver_resolve_freq() -> > clamp_val(target_freq, policy->min, policy->max) (1) > > by subtracting CPU's Thermal Pressure (ThPr) signal from the original > CPU capacity `arch_scale_cpu_capacity()` (2). > > --- > > Isn't there a conceptional flaw in this design? Let's say we have a > big.Little system with two cpufreq cooling devices and a thermal zone > (something like Hikey 960). To create a ThPr scenario we have to run > stuff on the CPUs (e.g. hackbench (3)). > Eventually cpufreq_set_cur_state() [drivers/thermal/cpufreq_cooling.c] > will set thermal_pressure to `(2) - (2)*freq/policy->cpuinfo.max_freq` > and PELT will provide the ThPr signal via thermal_load_avg(). > But to create this scenario, the system will become overutilized > (system-wide data, if one CPU is overutilized, the whole system is) so > EAS is disabled (i.e. find_energy_efficient_cpu() and compute_emergy() > are not executed). Not always, it depends on thermal governor decision, workload and 'power actors' (in IPA naming convention). Then it depends when and how hard you clamp the CPUs. They (CPUs) don't have to be always overutilized, they might be even 50-70% utilized but the GPU reduced power budget by 2 Watts, so CPUs left with only 1W. Which is still OK for the CPUs, since they are only 'feeding' the GPU with new 'jobs'. > > I can see that there are episodes in which EAS is running and > thermal_load_avg() != 0 but those have to be when (3) has stopped and > you see the ThPr signal just decaying (no accruing of new ThPr). The > cpufreq cooling device can still issue cpufreq_set_cur_state() but only > with decreasing states. It is true for some CPU heavy workloads, when no other SoC components are involved like: GPU, DSP, ISP, encoders, etc. For other workloads when CPUs don't have to do a lot, but thermal pressure might be seen on them, this patch help. > > --- > > IMHO, a precise description of how you envision the system setup, > incorporating all participating subsystems, would be helpful here. True, I hope this description above would help to understand the scenario. > >> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> >> --- >> kernel/sched/fair.c | 17 ++++++++++++++--- >> 1 file changed, 14 insertions(+), 3 deletions(-) >> >> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c >> index 161b92aa1c79..1aeddecabc20 100644 >> --- a/kernel/sched/fair.c >> +++ b/kernel/sched/fair.c >> @@ -6527,6 +6527,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) >> struct cpumask *pd_mask = perf_domain_span(pd); >> unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); >> unsigned long max_util = 0, sum_util = 0; >> + unsigned long _cpu_cap = cpu_cap; >> int cpu; >> >> /* >> @@ -6558,14 +6559,24 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) >> cpu_util_next(cpu, p, -1) + task_util_est(p); >> } >> >> + /* >> + * Take the thermal pressure from non-idle CPUs. They have >> + * most up-to-date information. For idle CPUs thermal pressure >> + * signal is not updated so often. >> + */ >> + if (!idle_cpu(cpu)) >> + _cpu_cap = cpu_cap - thermal_load_avg(cpu_rq(cpu)); >> + > > This one is probably the result of the fact that cpufreq cooling device > sets the ThPr for all CPUs of the policy (Frequency Domain (FD) or > Performance Domain (PD)) but PELT updates are happening per-CPU. And > only !idle CPUs get the update in scheduler_tick(). > > Looks like thermal_pressure [per_cpu(thermal_pressure, cpu), > drivers/base/arch_topology.c] set by cpufreq_set_cur_state() is always > in sync with policy->max/cpuinfo_max_freq). > So for your use case this instantaneous `signal` is better than the PELT > one. It's precise (no decaying when frequency clamping is already gone) > and you avoid the per-cpu update issue. Yes, this code implementation tries to address those issues.
On Thu, 10 Jun 2021 at 10:42, Lukasz Luba <lukasz.luba@arm.com> wrote: > > > > On 6/10/21 8:59 AM, Vincent Guittot wrote: > > On Fri, 4 Jun 2021 at 10:10, Lukasz Luba <lukasz.luba@arm.com> wrote: > >> > >> Energy Aware Scheduling (EAS) needs to be able to predict the frequency > >> requests made by the SchedUtil governor to properly estimate energy used > >> in the future. It has to take into account CPUs utilization and forecast > >> Performance Domain (PD) frequency. There is a corner case when the max > >> allowed frequency might be reduced due to thermal. SchedUtil is aware of > >> that reduced frequency, so it should be taken into account also in EAS > >> estimations. > >> > >> SchedUtil, as a CPUFreq governor, knows the maximum allowed frequency of > >> a CPU, thanks to cpufreq_driver_resolve_freq() and internal clamping > >> to 'policy::max'. SchedUtil is responsible to respect that upper limit > >> while setting the frequency through CPUFreq drivers. This effective > >> frequency is stored internally in 'sugov_policy::next_freq' and EAS has > >> to predict that value. > >> > >> In the existing code the raw value of arch_scale_cpu_capacity() is used > >> for clamping the returned CPU utilization from effective_cpu_util(). > >> This patch fixes issue with too big single CPU utilization, by introducing > >> clamping to the allowed CPU capacity. The allowed CPU capacity is a CPU > >> capacity reduced by thermal pressure signal. We rely on this load avg > >> geometric series in similar way as other mechanisms in the scheduler. > >> > >> Thanks to knowledge about allowed CPU capacity, we don't get too big value > >> for a single CPU utilization, which is then added to the util sum. The > >> util sum is used as a source of information for estimating whole PD energy. > >> To avoid wrong energy estimation in EAS (due to capped frequency), make > >> sure that the calculation of util sum is aware of allowed CPU capacity. > >> > >> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> > >> --- > >> kernel/sched/fair.c | 17 ++++++++++++++--- > >> 1 file changed, 14 insertions(+), 3 deletions(-) > >> > >> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c > >> index 161b92aa1c79..1aeddecabc20 100644 > >> --- a/kernel/sched/fair.c > >> +++ b/kernel/sched/fair.c > >> @@ -6527,6 +6527,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > >> struct cpumask *pd_mask = perf_domain_span(pd); > >> unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); > >> unsigned long max_util = 0, sum_util = 0; > >> + unsigned long _cpu_cap = cpu_cap; > >> int cpu; > >> > >> /* > >> @@ -6558,14 +6559,24 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > >> cpu_util_next(cpu, p, -1) + task_util_est(p); > >> } > >> > >> + /* > >> + * Take the thermal pressure from non-idle CPUs. They have > >> + * most up-to-date information. For idle CPUs thermal pressure > >> + * signal is not updated so often. > > > > What do you mean by "not updated so often" ? Do you have a value ? > > > > Thermal pressure is updated at the same rate as other PELT values of > > an idle CPU. Why is it a problem there ? > > > > > For idle CPU the value is updated 'remotely' by some other CPU > running nohz_idle_balance(). That goes into > update_blocked_averages() if the flags and checks are OK inside > update_nohz_stats(). Sometimes this is not called > because other_have_blocked() returned false. It can happen for a long So i miss that you were in a loop and the below was called for each cpu and _cpu_cap was overwritten + if (!idle_cpu(cpu)) + _cpu_cap = cpu_cap - thermal_load_avg(cpu_rq(cpu)); But that also means that if the 1st cpus of the pd are idle, they will use original capacity whereas the other ones will remove the thermal pressure. Isn't this a problem ? You don't use the same capacity for all cpus in the performance domain regarding the thermal pressure? > idle CPU, which all signals in that function has 0 [1]. > > This will cause that we don't check what is a new value stored by > thermal cpufreq_cooling for the thermal pressure [2]. We should feed > that value into the 'signal' machinery inside the > __update_blocked_others() [3]. Unfortunately, in a corner case there's > a flag (rq->has_blocked_load) which blocks the check of a > raw thermal value and prevents feeding it into thermal pressure signal > (since it's a long idle CPU, there is no load) [4]. > > It has implication on this patch, because I cannot e.g. take first > CPU from the PD mask and blindly check it's thermal pressure, > because it can be idle for a long time. I don't want to have two > loop, first just for taking the latest thermal pressure for the PD. > Thus, I want to re-use the existing loop to take the latest information > from non-idle CPU and pass use. > > Regards, > Lukasz > > > [1] https://elixir.bootlin.com/linux/latest/source/kernel/sched/fair.c#L7909 > [2] > https://elixir.bootlin.com/linux/latest/source/drivers/thermal/cpufreq_cooling.c#L494 > [3] https://elixir.bootlin.com/linux/latest/source/kernel/sched/fair.c#L7958 > [4] https://elixir.bootlin.com/linux/latest/source/kernel/sched/fair.c#L8433
On 6/10/21 10:11 AM, Vincent Guittot wrote: > On Thu, 10 Jun 2021 at 10:42, Lukasz Luba <lukasz.luba@arm.com> wrote: >> >> >> >> On 6/10/21 8:59 AM, Vincent Guittot wrote: >>> On Fri, 4 Jun 2021 at 10:10, Lukasz Luba <lukasz.luba@arm.com> wrote: >>>> >>>> Energy Aware Scheduling (EAS) needs to be able to predict the frequency >>>> requests made by the SchedUtil governor to properly estimate energy used >>>> in the future. It has to take into account CPUs utilization and forecast >>>> Performance Domain (PD) frequency. There is a corner case when the max >>>> allowed frequency might be reduced due to thermal. SchedUtil is aware of >>>> that reduced frequency, so it should be taken into account also in EAS >>>> estimations. >>>> >>>> SchedUtil, as a CPUFreq governor, knows the maximum allowed frequency of >>>> a CPU, thanks to cpufreq_driver_resolve_freq() and internal clamping >>>> to 'policy::max'. SchedUtil is responsible to respect that upper limit >>>> while setting the frequency through CPUFreq drivers. This effective >>>> frequency is stored internally in 'sugov_policy::next_freq' and EAS has >>>> to predict that value. >>>> >>>> In the existing code the raw value of arch_scale_cpu_capacity() is used >>>> for clamping the returned CPU utilization from effective_cpu_util(). >>>> This patch fixes issue with too big single CPU utilization, by introducing >>>> clamping to the allowed CPU capacity. The allowed CPU capacity is a CPU >>>> capacity reduced by thermal pressure signal. We rely on this load avg >>>> geometric series in similar way as other mechanisms in the scheduler. >>>> >>>> Thanks to knowledge about allowed CPU capacity, we don't get too big value >>>> for a single CPU utilization, which is then added to the util sum. The >>>> util sum is used as a source of information for estimating whole PD energy. >>>> To avoid wrong energy estimation in EAS (due to capped frequency), make >>>> sure that the calculation of util sum is aware of allowed CPU capacity. >>>> >>>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> >>>> --- >>>> kernel/sched/fair.c | 17 ++++++++++++++--- >>>> 1 file changed, 14 insertions(+), 3 deletions(-) >>>> >>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c >>>> index 161b92aa1c79..1aeddecabc20 100644 >>>> --- a/kernel/sched/fair.c >>>> +++ b/kernel/sched/fair.c >>>> @@ -6527,6 +6527,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) >>>> struct cpumask *pd_mask = perf_domain_span(pd); >>>> unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); >>>> unsigned long max_util = 0, sum_util = 0; >>>> + unsigned long _cpu_cap = cpu_cap; >>>> int cpu; >>>> >>>> /* >>>> @@ -6558,14 +6559,24 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) >>>> cpu_util_next(cpu, p, -1) + task_util_est(p); >>>> } >>>> >>>> + /* >>>> + * Take the thermal pressure from non-idle CPUs. They have >>>> + * most up-to-date information. For idle CPUs thermal pressure >>>> + * signal is not updated so often. >>> >>> What do you mean by "not updated so often" ? Do you have a value ? >>> >>> Thermal pressure is updated at the same rate as other PELT values of >>> an idle CPU. Why is it a problem there ? >>> >> >> >> For idle CPU the value is updated 'remotely' by some other CPU >> running nohz_idle_balance(). That goes into >> update_blocked_averages() if the flags and checks are OK inside >> update_nohz_stats(). Sometimes this is not called >> because other_have_blocked() returned false. It can happen for a long > > So i miss that you were in a loop and the below was called for each > cpu and _cpu_cap was overwritten > > + if (!idle_cpu(cpu)) > + _cpu_cap = cpu_cap - thermal_load_avg(cpu_rq(cpu)); > > But that also means that if the 1st cpus of the pd are idle, they will > use original capacity whereas the other ones will remove the thermal > pressure. Isn't this a problem ? You don't use the same capacity for > all cpus in the performance domain regarding the thermal pressure? True, but in the experiments for idle CPUs I haven't observed that they still have some big util (bigger than _cpu_cap). It decayed already, so it's not a problem for idle CPUs. Although, it might be my test case which didn't trigger something. Is it worth to add the loop above this one, to be 100% sure and get a thermal pressure signal from some running CPU? Then apply the same value always inside the 2nd loop?
On Thu, 10 Jun 2021 at 11:36, Lukasz Luba <lukasz.luba@arm.com> wrote: > > > > On 6/10/21 10:11 AM, Vincent Guittot wrote: > > On Thu, 10 Jun 2021 at 10:42, Lukasz Luba <lukasz.luba@arm.com> wrote: > >> > >> > >> > >> On 6/10/21 8:59 AM, Vincent Guittot wrote: > >>> On Fri, 4 Jun 2021 at 10:10, Lukasz Luba <lukasz.luba@arm.com> wrote: > >>>> > >>>> Energy Aware Scheduling (EAS) needs to be able to predict the frequency > >>>> requests made by the SchedUtil governor to properly estimate energy used > >>>> in the future. It has to take into account CPUs utilization and forecast > >>>> Performance Domain (PD) frequency. There is a corner case when the max > >>>> allowed frequency might be reduced due to thermal. SchedUtil is aware of > >>>> that reduced frequency, so it should be taken into account also in EAS > >>>> estimations. > >>>> > >>>> SchedUtil, as a CPUFreq governor, knows the maximum allowed frequency of > >>>> a CPU, thanks to cpufreq_driver_resolve_freq() and internal clamping > >>>> to 'policy::max'. SchedUtil is responsible to respect that upper limit > >>>> while setting the frequency through CPUFreq drivers. This effective > >>>> frequency is stored internally in 'sugov_policy::next_freq' and EAS has > >>>> to predict that value. > >>>> > >>>> In the existing code the raw value of arch_scale_cpu_capacity() is used > >>>> for clamping the returned CPU utilization from effective_cpu_util(). > >>>> This patch fixes issue with too big single CPU utilization, by introducing > >>>> clamping to the allowed CPU capacity. The allowed CPU capacity is a CPU > >>>> capacity reduced by thermal pressure signal. We rely on this load avg > >>>> geometric series in similar way as other mechanisms in the scheduler. > >>>> > >>>> Thanks to knowledge about allowed CPU capacity, we don't get too big value > >>>> for a single CPU utilization, which is then added to the util sum. The > >>>> util sum is used as a source of information for estimating whole PD energy. > >>>> To avoid wrong energy estimation in EAS (due to capped frequency), make > >>>> sure that the calculation of util sum is aware of allowed CPU capacity. > >>>> > >>>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> > >>>> --- > >>>> kernel/sched/fair.c | 17 ++++++++++++++--- > >>>> 1 file changed, 14 insertions(+), 3 deletions(-) > >>>> > >>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c > >>>> index 161b92aa1c79..1aeddecabc20 100644 > >>>> --- a/kernel/sched/fair.c > >>>> +++ b/kernel/sched/fair.c > >>>> @@ -6527,6 +6527,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > >>>> struct cpumask *pd_mask = perf_domain_span(pd); > >>>> unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); > >>>> unsigned long max_util = 0, sum_util = 0; > >>>> + unsigned long _cpu_cap = cpu_cap; > >>>> int cpu; > >>>> > >>>> /* > >>>> @@ -6558,14 +6559,24 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > >>>> cpu_util_next(cpu, p, -1) + task_util_est(p); > >>>> } > >>>> > >>>> + /* > >>>> + * Take the thermal pressure from non-idle CPUs. They have > >>>> + * most up-to-date information. For idle CPUs thermal pressure > >>>> + * signal is not updated so often. > >>> > >>> What do you mean by "not updated so often" ? Do you have a value ? > >>> > >>> Thermal pressure is updated at the same rate as other PELT values of > >>> an idle CPU. Why is it a problem there ? > >>> > >> > >> > >> For idle CPU the value is updated 'remotely' by some other CPU > >> running nohz_idle_balance(). That goes into > >> update_blocked_averages() if the flags and checks are OK inside > >> update_nohz_stats(). Sometimes this is not called > >> because other_have_blocked() returned false. It can happen for a long > > > > So i miss that you were in a loop and the below was called for each > > cpu and _cpu_cap was overwritten > > > > + if (!idle_cpu(cpu)) > > + _cpu_cap = cpu_cap - thermal_load_avg(cpu_rq(cpu)); > > > > But that also means that if the 1st cpus of the pd are idle, they will > > use original capacity whereas the other ones will remove the thermal > > pressure. Isn't this a problem ? You don't use the same capacity for > > all cpus in the performance domain regarding the thermal pressure? > > True, but in the experiments for idle CPUs I haven't > observed that they still have some big util (bigger than _cpu_cap). > It decayed already, so it's not a problem for idle CPUs. But it's a problem because there is a random behavior : some idle cpu will use original capacity whereas others will use the capped value set by non idle CPUs. You must have consistent behavior across all idle cpus. Then, if it's not a problem why adding the if (!idle_cpu(cpu)) > > Although, it might be my test case which didn't trigger something. > Is it worth to add the loop above this one, to be 100% sure and > get a thermal pressure signal from some running CPU? > Then apply the same value always inside the 2nd loop?
On Thu, 10 Jun 2021 at 11:36, Lukasz Luba <lukasz.luba@arm.com> wrote: > > > > On 6/10/21 10:11 AM, Vincent Guittot wrote: > > On Thu, 10 Jun 2021 at 10:42, Lukasz Luba <lukasz.luba@arm.com> wrote: > >> > >> > >> > >> On 6/10/21 8:59 AM, Vincent Guittot wrote: > >>> On Fri, 4 Jun 2021 at 10:10, Lukasz Luba <lukasz.luba@arm.com> wrote: > >>>> > >>>> Energy Aware Scheduling (EAS) needs to be able to predict the frequency > >>>> requests made by the SchedUtil governor to properly estimate energy used > >>>> in the future. It has to take into account CPUs utilization and forecast > >>>> Performance Domain (PD) frequency. There is a corner case when the max > >>>> allowed frequency might be reduced due to thermal. SchedUtil is aware of > >>>> that reduced frequency, so it should be taken into account also in EAS > >>>> estimations. > >>>> > >>>> SchedUtil, as a CPUFreq governor, knows the maximum allowed frequency of > >>>> a CPU, thanks to cpufreq_driver_resolve_freq() and internal clamping > >>>> to 'policy::max'. SchedUtil is responsible to respect that upper limit > >>>> while setting the frequency through CPUFreq drivers. This effective > >>>> frequency is stored internally in 'sugov_policy::next_freq' and EAS has > >>>> to predict that value. > >>>> > >>>> In the existing code the raw value of arch_scale_cpu_capacity() is used > >>>> for clamping the returned CPU utilization from effective_cpu_util(). > >>>> This patch fixes issue with too big single CPU utilization, by introducing > >>>> clamping to the allowed CPU capacity. The allowed CPU capacity is a CPU > >>>> capacity reduced by thermal pressure signal. We rely on this load avg > >>>> geometric series in similar way as other mechanisms in the scheduler. > >>>> > >>>> Thanks to knowledge about allowed CPU capacity, we don't get too big value > >>>> for a single CPU utilization, which is then added to the util sum. The > >>>> util sum is used as a source of information for estimating whole PD energy. > >>>> To avoid wrong energy estimation in EAS (due to capped frequency), make > >>>> sure that the calculation of util sum is aware of allowed CPU capacity. > >>>> > >>>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> > >>>> --- > >>>> kernel/sched/fair.c | 17 ++++++++++++++--- > >>>> 1 file changed, 14 insertions(+), 3 deletions(-) > >>>> > >>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c > >>>> index 161b92aa1c79..1aeddecabc20 100644 > >>>> --- a/kernel/sched/fair.c > >>>> +++ b/kernel/sched/fair.c > >>>> @@ -6527,6 +6527,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > >>>> struct cpumask *pd_mask = perf_domain_span(pd); > >>>> unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); > >>>> unsigned long max_util = 0, sum_util = 0; > >>>> + unsigned long _cpu_cap = cpu_cap; > >>>> int cpu; > >>>> > >>>> /* > >>>> @@ -6558,14 +6559,24 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > >>>> cpu_util_next(cpu, p, -1) + task_util_est(p); > >>>> } > >>>> > >>>> + /* > >>>> + * Take the thermal pressure from non-idle CPUs. They have > >>>> + * most up-to-date information. For idle CPUs thermal pressure > >>>> + * signal is not updated so often. > >>> > >>> What do you mean by "not updated so often" ? Do you have a value ? > >>> > >>> Thermal pressure is updated at the same rate as other PELT values of > >>> an idle CPU. Why is it a problem there ? > >>> > >> > >> > >> For idle CPU the value is updated 'remotely' by some other CPU > >> running nohz_idle_balance(). That goes into > >> update_blocked_averages() if the flags and checks are OK inside > >> update_nohz_stats(). Sometimes this is not called > >> because other_have_blocked() returned false. It can happen for a long > > > > So i miss that you were in a loop and the below was called for each > > cpu and _cpu_cap was overwritten > > > > + if (!idle_cpu(cpu)) > > + _cpu_cap = cpu_cap - thermal_load_avg(cpu_rq(cpu)); > > > > But that also means that if the 1st cpus of the pd are idle, they will > > use original capacity whereas the other ones will remove the thermal > > pressure. Isn't this a problem ? You don't use the same capacity for > > all cpus in the performance domain regarding the thermal pressure? > > True, but in the experiments for idle CPUs I haven't > observed that they still have some big util (bigger than _cpu_cap). > It decayed already, so it's not a problem for idle CPUs. > > Although, it might be my test case which didn't trigger something. > Is it worth to add the loop above this one, to be 100% sure and > get a thermal pressure signal from some running CPU? > Then apply the same value always inside the 2nd loop? Either it's a problem and you must make sure to use the same capacity for all cpus of a PD Or it's not but in this case you don't need if (!idle_cpu(cpu))
On 6/10/21 10:41 AM, Vincent Guittot wrote: > On Thu, 10 Jun 2021 at 11:36, Lukasz Luba <lukasz.luba@arm.com> wrote: >> >> >> >> On 6/10/21 10:11 AM, Vincent Guittot wrote: >>> On Thu, 10 Jun 2021 at 10:42, Lukasz Luba <lukasz.luba@arm.com> wrote: >>>> >>>> >>>> >>>> On 6/10/21 8:59 AM, Vincent Guittot wrote: >>>>> On Fri, 4 Jun 2021 at 10:10, Lukasz Luba <lukasz.luba@arm.com> wrote: >>>>>> >>>>>> Energy Aware Scheduling (EAS) needs to be able to predict the frequency >>>>>> requests made by the SchedUtil governor to properly estimate energy used >>>>>> in the future. It has to take into account CPUs utilization and forecast >>>>>> Performance Domain (PD) frequency. There is a corner case when the max >>>>>> allowed frequency might be reduced due to thermal. SchedUtil is aware of >>>>>> that reduced frequency, so it should be taken into account also in EAS >>>>>> estimations. >>>>>> >>>>>> SchedUtil, as a CPUFreq governor, knows the maximum allowed frequency of >>>>>> a CPU, thanks to cpufreq_driver_resolve_freq() and internal clamping >>>>>> to 'policy::max'. SchedUtil is responsible to respect that upper limit >>>>>> while setting the frequency through CPUFreq drivers. This effective >>>>>> frequency is stored internally in 'sugov_policy::next_freq' and EAS has >>>>>> to predict that value. >>>>>> >>>>>> In the existing code the raw value of arch_scale_cpu_capacity() is used >>>>>> for clamping the returned CPU utilization from effective_cpu_util(). >>>>>> This patch fixes issue with too big single CPU utilization, by introducing >>>>>> clamping to the allowed CPU capacity. The allowed CPU capacity is a CPU >>>>>> capacity reduced by thermal pressure signal. We rely on this load avg >>>>>> geometric series in similar way as other mechanisms in the scheduler. >>>>>> >>>>>> Thanks to knowledge about allowed CPU capacity, we don't get too big value >>>>>> for a single CPU utilization, which is then added to the util sum. The >>>>>> util sum is used as a source of information for estimating whole PD energy. >>>>>> To avoid wrong energy estimation in EAS (due to capped frequency), make >>>>>> sure that the calculation of util sum is aware of allowed CPU capacity. >>>>>> >>>>>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> >>>>>> --- >>>>>> kernel/sched/fair.c | 17 ++++++++++++++--- >>>>>> 1 file changed, 14 insertions(+), 3 deletions(-) >>>>>> >>>>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c >>>>>> index 161b92aa1c79..1aeddecabc20 100644 >>>>>> --- a/kernel/sched/fair.c >>>>>> +++ b/kernel/sched/fair.c >>>>>> @@ -6527,6 +6527,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) >>>>>> struct cpumask *pd_mask = perf_domain_span(pd); >>>>>> unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); >>>>>> unsigned long max_util = 0, sum_util = 0; >>>>>> + unsigned long _cpu_cap = cpu_cap; >>>>>> int cpu; >>>>>> >>>>>> /* >>>>>> @@ -6558,14 +6559,24 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) >>>>>> cpu_util_next(cpu, p, -1) + task_util_est(p); >>>>>> } >>>>>> >>>>>> + /* >>>>>> + * Take the thermal pressure from non-idle CPUs. They have >>>>>> + * most up-to-date information. For idle CPUs thermal pressure >>>>>> + * signal is not updated so often. >>>>> >>>>> What do you mean by "not updated so often" ? Do you have a value ? >>>>> >>>>> Thermal pressure is updated at the same rate as other PELT values of >>>>> an idle CPU. Why is it a problem there ? >>>>> >>>> >>>> >>>> For idle CPU the value is updated 'remotely' by some other CPU >>>> running nohz_idle_balance(). That goes into >>>> update_blocked_averages() if the flags and checks are OK inside >>>> update_nohz_stats(). Sometimes this is not called >>>> because other_have_blocked() returned false. It can happen for a long >>> >>> So i miss that you were in a loop and the below was called for each >>> cpu and _cpu_cap was overwritten >>> >>> + if (!idle_cpu(cpu)) >>> + _cpu_cap = cpu_cap - thermal_load_avg(cpu_rq(cpu)); >>> >>> But that also means that if the 1st cpus of the pd are idle, they will >>> use original capacity whereas the other ones will remove the thermal >>> pressure. Isn't this a problem ? You don't use the same capacity for >>> all cpus in the performance domain regarding the thermal pressure? >> >> True, but in the experiments for idle CPUs I haven't >> observed that they still have some big util (bigger than _cpu_cap). >> It decayed already, so it's not a problem for idle CPUs. > > But it's a problem because there is a random behavior : some idle cpu > will use original capacity whereas others will use the capped value > set by non idle CPUs. You must have consistent behavior across all > idle cpus. > > Then, if it's not a problem why adding the if (!idle_cpu(cpu)) To capture the signal value from a running CPU, which then I pass into the em_cpu_energy() in path 2/2. My apologies for confusion, this can be just local variable for patch 1/2. I can create the _cpu_cap as local variable inside this loop, just for this patch. Then in patch 2/2 I will remove it and define above the loop, to be available for the call to em_cpu_energy().
On 10/06/2021 11:04, Lukasz Luba wrote: > > > On 6/10/21 9:42 AM, Dietmar Eggemann wrote: > > [snip] > >> >> So essentially what you want to do is: >> >> Make EAS aware of the frequency clamping schedutil can be faced with: >> >> get_next_freq() -> cpufreq_driver_resolve_freq() -> >> clamp_val(target_freq, policy->min, policy->max) (1) >> >> by subtracting CPU's Thermal Pressure (ThPr) signal from the original >> CPU capacity `arch_scale_cpu_capacity()` (2). >> >> --- >> >> Isn't there a conceptional flaw in this design? Let's say we have a >> big.Little system with two cpufreq cooling devices and a thermal zone >> (something like Hikey 960). To create a ThPr scenario we have to run >> stuff on the CPUs (e.g. hackbench (3)). >> Eventually cpufreq_set_cur_state() [drivers/thermal/cpufreq_cooling.c] >> will set thermal_pressure to `(2) - (2)*freq/policy->cpuinfo.max_freq` >> and PELT will provide the ThPr signal via thermal_load_avg(). >> But to create this scenario, the system will become overutilized >> (system-wide data, if one CPU is overutilized, the whole system is) so >> EAS is disabled (i.e. find_energy_efficient_cpu() and compute_emergy() >> are not executed). > > Not always, it depends on thermal governor decision, workload and > 'power actors' (in IPA naming convention). Then it depends when and how > hard you clamp the CPUs. They (CPUs) don't have to be always > overutilized, they might be even 50-70% utilized but the GPU reduced > power budget by 2 Watts, so CPUs left with only 1W. Which is still OK > for the CPUs, since they are only 'feeding' the GPU with new 'jobs'. All this pretty much confines the usefulness of you proposed change. A precise description of it with the patches is necessary to allow people to start from there while exploring your patches. >> I can see that there are episodes in which EAS is running and >> thermal_load_avg() != 0 but those have to be when (3) has stopped and >> you see the ThPr signal just decaying (no accruing of new ThPr). The >> cpufreq cooling device can still issue cpufreq_set_cur_state() but only >> with decreasing states. > > It is true for some CPU heavy workloads, when no other SoC components > are involved like: GPU, DSP, ISP, encoders, etc. For other workloads > when CPUs don't have to do a lot, but thermal pressure might be seen on > them, this patch help. > >> >> --- >> >> IMHO, a precise description of how you envision the system setup, >> incorporating all participating subsystems, would be helpful here. > > True, I hope this description above would help to understand the > scenario. This description belongs in the patch header. The scenario in which your functionality would improve things has to be clear. I'm sure that not everybody looking at this patches is immediately aware on how IPA setups work and which specific setup you have in mind here. >>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> >>> --- >>> kernel/sched/fair.c | 17 ++++++++++++++--- >>> 1 file changed, 14 insertions(+), 3 deletions(-) >>> >>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c >>> index 161b92aa1c79..1aeddecabc20 100644 >>> --- a/kernel/sched/fair.c >>> +++ b/kernel/sched/fair.c >>> @@ -6527,6 +6527,7 @@ compute_energy(struct task_struct *p, int >>> dst_cpu, struct perf_domain *pd) >>> struct cpumask *pd_mask = perf_domain_span(pd); >>> unsigned long cpu_cap = >>> arch_scale_cpu_capacity(cpumask_first(pd_mask)); >>> unsigned long max_util = 0, sum_util = 0; >>> + unsigned long _cpu_cap = cpu_cap; >>> int cpu; >>> /* >>> @@ -6558,14 +6559,24 @@ compute_energy(struct task_struct *p, int >>> dst_cpu, struct perf_domain *pd) >>> cpu_util_next(cpu, p, -1) + task_util_est(p); >>> } >>> + /* >>> + * Take the thermal pressure from non-idle CPUs. They have >>> + * most up-to-date information. For idle CPUs thermal pressure >>> + * signal is not updated so often. >>> + */ >>> + if (!idle_cpu(cpu)) >>> + _cpu_cap = cpu_cap - thermal_load_avg(cpu_rq(cpu)); >>> + >> >> This one is probably the result of the fact that cpufreq cooling device >> sets the ThPr for all CPUs of the policy (Frequency Domain (FD) or >> Performance Domain (PD)) but PELT updates are happening per-CPU. And >> only !idle CPUs get the update in scheduler_tick(). >> >> Looks like thermal_pressure [per_cpu(thermal_pressure, cpu), >> drivers/base/arch_topology.c] set by cpufreq_set_cur_state() is always >> in sync with policy->max/cpuinfo_max_freq). >> So for your use case this instantaneous `signal` is better than the PELT >> one. It's precise (no decaying when frequency clamping is already gone) >> and you avoid the per-cpu update issue. > > Yes, this code implementation tries to address those issues. The point I was making here is: why using the PELT signal thermal_load_avg() and not per_cpu(thermal_pressure, cpu) directly, given the fact that the latter perfectly represents the frequency clamping?
On 6/10/21 11:07 AM, Dietmar Eggemann wrote: > On 10/06/2021 11:04, Lukasz Luba wrote: >> [snip] >> Not always, it depends on thermal governor decision, workload and >> 'power actors' (in IPA naming convention). Then it depends when and how >> hard you clamp the CPUs. They (CPUs) don't have to be always >> overutilized, they might be even 50-70% utilized but the GPU reduced >> power budget by 2 Watts, so CPUs left with only 1W. Which is still OK >> for the CPUs, since they are only 'feeding' the GPU with new 'jobs'. > > All this pretty much confines the usefulness of you proposed change. A > precise description of it with the patches is necessary to allow people > to start from there while exploring your patches. OK, I see your point. [snip] >> True, I hope this description above would help to understand the >> scenario. > > This description belongs in the patch header. The scenario in which your > functionality would improve things has to be clear. > I'm sure that not everybody looking at this patches is immediately aware > on how IPA setups work and which specific setup you have in mind here. Agree. I will add this description into the patch header for v3. [snip] >> >> Yes, this code implementation tries to address those issues. > > The point I was making here is: why using the PELT signal > thermal_load_avg() and not per_cpu(thermal_pressure, cpu) directly, > given the fact that the latter perfectly represents the frequency clamping? > Good question. I wanted to be aligned with other parts in the fair.c like cpu_capacity() and all it's users. The CPU capacity is reduced by RT, DL, IRQ and thermal load avg, not the 'raw' value from the per-cpu variable. TBH I cannot recall what was the argument back then when thermal pressure geometric series was introduced. Maybe to have a better control how fast it raises and decays so other mechanisms in the scheduler will see the change in thermal as not so sharp... (?) Vincent do you remember the motivation to have geometric series in thermal pressure and not use just the 'raw' value from per-cpu?
On Thu, 10 Jun 2021 at 12:37, Lukasz Luba <lukasz.luba@arm.com> wrote: > > > > On 6/10/21 11:07 AM, Dietmar Eggemann wrote: > > On 10/06/2021 11:04, Lukasz Luba wrote: > >> > > [snip] > > >> Not always, it depends on thermal governor decision, workload and > >> 'power actors' (in IPA naming convention). Then it depends when and how > >> hard you clamp the CPUs. They (CPUs) don't have to be always > >> overutilized, they might be even 50-70% utilized but the GPU reduced > >> power budget by 2 Watts, so CPUs left with only 1W. Which is still OK > >> for the CPUs, since they are only 'feeding' the GPU with new 'jobs'. > > > > All this pretty much confines the usefulness of you proposed change. A > > precise description of it with the patches is necessary to allow people > > to start from there while exploring your patches. > > OK, I see your point. > > [snip] > > >> True, I hope this description above would help to understand the > >> scenario. > > > > This description belongs in the patch header. The scenario in which your > > functionality would improve things has to be clear. > > I'm sure that not everybody looking at this patches is immediately aware > > on how IPA setups work and which specific setup you have in mind here. > > Agree. I will add this description into the patch header for v3. > > [snip] > > >> > >> Yes, this code implementation tries to address those issues. > > > > The point I was making here is: why using the PELT signal > > thermal_load_avg() and not per_cpu(thermal_pressure, cpu) directly, > > given the fact that the latter perfectly represents the frequency clamping? > > > > Good question. I wanted to be aligned with other parts in the fair.c > like cpu_capacity() and all it's users. The CPU capacity is reduced by > RT, DL, IRQ and thermal load avg, not the 'raw' value from the > per-cpu variable. > > TBH I cannot recall what was the argument back then > when thermal pressure geometric series was introduced. > Maybe to have a better control how fast it raises and decays > so other mechanisms in the scheduler will see the change in thermal > as not so sharp... (?) > > > Vincent do you remember the motivation to have geometric series > in thermal pressure and not use just the 'raw' value from per-cpu? In order to have thermal pressure synced with other metrics used by the scheduler like util/rt/dl_avg. As an example, when thermal pressure will decrease because thermal capping is removed, the utilization will increase at the same pace as thermal will decrease and it will not create some fake spare cycle. util_avg is the average expected utilization of the cpu, thermal pressure reflects the average stolen capacity for the same averaging time scale but this can be the result of a toggle between several OPP Using current capping is quite volatile to make a decision as it might have changed by the time you apply your decision.
On 6/10/21 1:19 PM, Vincent Guittot wrote: > On Thu, 10 Jun 2021 at 12:37, Lukasz Luba <lukasz.luba@arm.com> wrote: >> >> >> >> On 6/10/21 11:07 AM, Dietmar Eggemann wrote: >>> On 10/06/2021 11:04, Lukasz Luba wrote: >>>> >> >> [snip] >> >>>> Not always, it depends on thermal governor decision, workload and >>>> 'power actors' (in IPA naming convention). Then it depends when and how >>>> hard you clamp the CPUs. They (CPUs) don't have to be always >>>> overutilized, they might be even 50-70% utilized but the GPU reduced >>>> power budget by 2 Watts, so CPUs left with only 1W. Which is still OK >>>> for the CPUs, since they are only 'feeding' the GPU with new 'jobs'. >>> >>> All this pretty much confines the usefulness of you proposed change. A >>> precise description of it with the patches is necessary to allow people >>> to start from there while exploring your patches. >> >> OK, I see your point. >> >> [snip] >> >>>> True, I hope this description above would help to understand the >>>> scenario. >>> >>> This description belongs in the patch header. The scenario in which your >>> functionality would improve things has to be clear. >>> I'm sure that not everybody looking at this patches is immediately aware >>> on how IPA setups work and which specific setup you have in mind here. >> >> Agree. I will add this description into the patch header for v3. >> >> [snip] >> >>>> >>>> Yes, this code implementation tries to address those issues. >>> >>> The point I was making here is: why using the PELT signal >>> thermal_load_avg() and not per_cpu(thermal_pressure, cpu) directly, >>> given the fact that the latter perfectly represents the frequency clamping? >>> >> >> Good question. I wanted to be aligned with other parts in the fair.c >> like cpu_capacity() and all it's users. The CPU capacity is reduced by >> RT, DL, IRQ and thermal load avg, not the 'raw' value from the >> per-cpu variable. >> >> TBH I cannot recall what was the argument back then >> when thermal pressure geometric series was introduced. >> Maybe to have a better control how fast it raises and decays >> so other mechanisms in the scheduler will see the change in thermal >> as not so sharp... (?) >> >> >> Vincent do you remember the motivation to have geometric series >> in thermal pressure and not use just the 'raw' value from per-cpu? > > In order to have thermal pressure synced with other metrics used by > the scheduler like util/rt/dl_avg. As an example, when thermal > pressure will decrease because thermal capping is removed, the > utilization will increase at the same pace as thermal will decrease > and it will not create some fake spare cycle. util_avg is the average > expected utilization of the cpu, thermal pressure reflects the average > stolen capacity for the same averaging time scale but this can be the > result of a toggle between several OPP > > Using current capping is quite volatile to make a decision as it might > have changed by the time you apply your decision. > So for this scenario, where we want to just align EAS with SchedUtil frequency decision, which is instantaneous and has 'raw' value of capping from policy->max, shouldn't we use: thermal_pressure = arch_scale_thermal_pressure(cpu_id) instead of geometric series thermal_pressure signal? This would avoid the hassling with idle CPUs and not updated thermal signal.
On Thu, 10 Jun 2021 at 14:30, Lukasz Luba <lukasz.luba@arm.com> wrote: > > > > On 6/10/21 1:19 PM, Vincent Guittot wrote: > > On Thu, 10 Jun 2021 at 12:37, Lukasz Luba <lukasz.luba@arm.com> wrote: > >> > >> > >> > >> On 6/10/21 11:07 AM, Dietmar Eggemann wrote: > >>> On 10/06/2021 11:04, Lukasz Luba wrote: > >>>> > >> > >> [snip] > >> > >>>> Not always, it depends on thermal governor decision, workload and > >>>> 'power actors' (in IPA naming convention). Then it depends when and how > >>>> hard you clamp the CPUs. They (CPUs) don't have to be always > >>>> overutilized, they might be even 50-70% utilized but the GPU reduced > >>>> power budget by 2 Watts, so CPUs left with only 1W. Which is still OK > >>>> for the CPUs, since they are only 'feeding' the GPU with new 'jobs'. > >>> > >>> All this pretty much confines the usefulness of you proposed change. A > >>> precise description of it with the patches is necessary to allow people > >>> to start from there while exploring your patches. > >> > >> OK, I see your point. > >> > >> [snip] > >> > >>>> True, I hope this description above would help to understand the > >>>> scenario. > >>> > >>> This description belongs in the patch header. The scenario in which your > >>> functionality would improve things has to be clear. > >>> I'm sure that not everybody looking at this patches is immediately aware > >>> on how IPA setups work and which specific setup you have in mind here. > >> > >> Agree. I will add this description into the patch header for v3. > >> > >> [snip] > >> > >>>> > >>>> Yes, this code implementation tries to address those issues. > >>> > >>> The point I was making here is: why using the PELT signal > >>> thermal_load_avg() and not per_cpu(thermal_pressure, cpu) directly, > >>> given the fact that the latter perfectly represents the frequency clamping? > >>> > >> > >> Good question. I wanted to be aligned with other parts in the fair.c > >> like cpu_capacity() and all it's users. The CPU capacity is reduced by > >> RT, DL, IRQ and thermal load avg, not the 'raw' value from the > >> per-cpu variable. > >> > >> TBH I cannot recall what was the argument back then > >> when thermal pressure geometric series was introduced. > >> Maybe to have a better control how fast it raises and decays > >> so other mechanisms in the scheduler will see the change in thermal > >> as not so sharp... (?) > >> > >> > >> Vincent do you remember the motivation to have geometric series > >> in thermal pressure and not use just the 'raw' value from per-cpu? > > > > In order to have thermal pressure synced with other metrics used by > > the scheduler like util/rt/dl_avg. As an example, when thermal > > pressure will decrease because thermal capping is removed, the > > utilization will increase at the same pace as thermal will decrease > > and it will not create some fake spare cycle. util_avg is the average > > expected utilization of the cpu, thermal pressure reflects the average > > stolen capacity for the same averaging time scale but this can be the > > result of a toggle between several OPP > > > > Using current capping is quite volatile to make a decision as it might > > have changed by the time you apply your decision. > > > > So for this scenario, where we want to just align EAS with SchedUtil > frequency decision, which is instantaneous and has 'raw' value > of capping from policy->max, shouldn't we use: > > thermal_pressure = arch_scale_thermal_pressure(cpu_id) Yes you should probably use arch_scale_thermal_pressure(cpu) instead of thermal_load_avg(rq) in this case > > instead of geometric series thermal_pressure signal? > > This would avoid the hassling with idle CPUs and not updated > thermal signal.
On 6/10/21 1:40 PM, Vincent Guittot wrote: > On Thu, 10 Jun 2021 at 14:30, Lukasz Luba <lukasz.luba@arm.com> wrote: [snip] >> >> So for this scenario, where we want to just align EAS with SchedUtil >> frequency decision, which is instantaneous and has 'raw' value >> of capping from policy->max, shouldn't we use: >> >> thermal_pressure = arch_scale_thermal_pressure(cpu_id) > > Yes you should probably use arch_scale_thermal_pressure(cpu) instead > of thermal_load_avg(rq) in this case > Thank you Vincent for valuable opinions! I will rewrite it and experiment with a new approach, then send a v3. Regards, Lukasz
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 161b92aa1c79..1aeddecabc20 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -6527,6 +6527,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) struct cpumask *pd_mask = perf_domain_span(pd); unsigned long cpu_cap = arch_scale_cpu_capacity(cpumask_first(pd_mask)); unsigned long max_util = 0, sum_util = 0; + unsigned long _cpu_cap = cpu_cap; int cpu; /* @@ -6558,14 +6559,24 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) cpu_util_next(cpu, p, -1) + task_util_est(p); } + /* + * Take the thermal pressure from non-idle CPUs. They have + * most up-to-date information. For idle CPUs thermal pressure + * signal is not updated so often. + */ + if (!idle_cpu(cpu)) + _cpu_cap = cpu_cap - thermal_load_avg(cpu_rq(cpu)); + /* * Busy time computation: utilization clamping is not * required since the ratio (sum_util / cpu_capacity) * is already enough to scale the EM reported power * consumption at the (eventually clamped) cpu_capacity. */ - sum_util += effective_cpu_util(cpu, util_running, cpu_cap, - ENERGY_UTIL, NULL); + cpu_util = effective_cpu_util(cpu, util_running, cpu_cap, + ENERGY_UTIL, NULL); + + sum_util += min(cpu_util, _cpu_cap); /* * Performance domain frequency: utilization clamping @@ -6576,7 +6587,7 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) */ cpu_util = effective_cpu_util(cpu, util_freq, cpu_cap, FREQUENCY_UTIL, tsk); - max_util = max(max_util, cpu_util); + max_util = max(max_util, min(cpu_util, _cpu_cap)); } return em_cpu_energy(pd->em_pd, max_util, sum_util);
Energy Aware Scheduling (EAS) needs to be able to predict the frequency requests made by the SchedUtil governor to properly estimate energy used in the future. It has to take into account CPUs utilization and forecast Performance Domain (PD) frequency. There is a corner case when the max allowed frequency might be reduced due to thermal. SchedUtil is aware of that reduced frequency, so it should be taken into account also in EAS estimations. SchedUtil, as a CPUFreq governor, knows the maximum allowed frequency of a CPU, thanks to cpufreq_driver_resolve_freq() and internal clamping to 'policy::max'. SchedUtil is responsible to respect that upper limit while setting the frequency through CPUFreq drivers. This effective frequency is stored internally in 'sugov_policy::next_freq' and EAS has to predict that value. In the existing code the raw value of arch_scale_cpu_capacity() is used for clamping the returned CPU utilization from effective_cpu_util(). This patch fixes issue with too big single CPU utilization, by introducing clamping to the allowed CPU capacity. The allowed CPU capacity is a CPU capacity reduced by thermal pressure signal. We rely on this load avg geometric series in similar way as other mechanisms in the scheduler. Thanks to knowledge about allowed CPU capacity, we don't get too big value for a single CPU utilization, which is then added to the util sum. The util sum is used as a source of information for estimating whole PD energy. To avoid wrong energy estimation in EAS (due to capped frequency), make sure that the calculation of util sum is aware of allowed CPU capacity. Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> --- kernel/sched/fair.c | 17 ++++++++++++++--- 1 file changed, 14 insertions(+), 3 deletions(-)