Hi all, The patch set aims to address a scenario for Energy Aware Scheduler, where we estimate and compare energy values and miss a more precised results. In some use cases estimations for two CPUs might give the same values for a given task and it's utilization. Those values would be different when we have a better precision and avoid this rounding error. Thus, the decision of choosing a CPU for a waking-up task might also be better. We have received this feedback from our partners. Address this rounding error issue and increase the precision of Energy Model em_perf_state::cost values. This change should not affect other subsystems in kernel: thermal IPA, PowerCap DTPM, etc, since they use em_perf_state::power field, which is not touched. It also doesn't trigger the need for updating all existing platforms to register EM and report power values in different scale. Regards, Lukasz Lukasz Luba (3): sched/fair: Prepare variables for increased precision of EAS estimated energy PM: EM: Make em_cpu_energy() able to return bigger values PM: EM: Increase energy calculation precision include/linux/energy_model.h | 11 +++++++---- kernel/power/energy_model.c | 3 ++- kernel/sched/fair.c | 13 +++++++------ 3 files changed, 16 insertions(+), 11 deletions(-) -- 2.17.1
The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up task. It probes many possibilities and compares the estimated energy values for different scenarios. For calculating those energy values it relies on Energy Model (EM) data and em_cpu_energy(). The precision which is used in EM data is in milli-Watts (or abstract scale), which sometimes is not sufficient. In some cases it might happen that two CPUs from different Performance Domains (PDs) get the same calculated value for a given task placement, but in more precised scale, they might differ. This rounding error has to be addressed. This patch prepares EAS code for better precision in the coming EM improvements. Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> --- kernel/sched/fair.c | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-) diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 7b8990fd4896..b517c9e79768 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -6582,7 +6582,7 @@ static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) * to compute what would be the energy if we decided to actually migrate that * task. */ -static long +static u64 compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) { struct cpumask *pd_mask = perf_domain_span(pd); @@ -6689,12 +6689,13 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) */ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) { - unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX; struct root_domain *rd = cpu_rq(smp_processor_id())->rd; + u64 prev_delta = ULLONG_MAX, best_delta = ULLONG_MAX; int cpu, best_energy_cpu = prev_cpu, target = -1; - unsigned long cpu_cap, util, base_energy = 0; + unsigned long cpu_cap, util; struct sched_domain *sd; struct perf_domain *pd; + u64 base_energy = 0; rcu_read_lock(); pd = rcu_dereference(rd->pd); @@ -6718,9 +6719,9 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) goto unlock; for (; pd; pd = pd->next) { - unsigned long cur_delta, spare_cap, max_spare_cap = 0; + unsigned long spare_cap, max_spare_cap = 0; bool compute_prev_delta = false; - unsigned long base_energy_pd; + u64 base_energy_pd, cur_delta; int max_spare_cap_cpu = -1; for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) { @@ -6790,7 +6791,7 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) * Pick the best CPU if prev_cpu cannot be used, or if it saves at * least 6% of the energy used by prev_cpu. */ - if ((prev_delta == ULONG_MAX) || + if ((prev_delta == ULLONG_MAX) || (prev_delta - best_delta) > ((prev_delta + base_energy) >> 4)) target = best_energy_cpu; -- 2.17.1
The Energy Model (EM) em_cpu_energy() is responsible for providing good estimation regarding CPUs energy. It contains proper data structures which are then used during calculation. The values stored in there are in milli-Watts precision (or in abstract scale) smaller that 0xffff, which use sufficient unsigned long even on 32-bit machines. There are scenarios where we would like to provide calculated estimations in a better precision and the values might be 1000 times bigger. This patch makes possible to use quite big values for also 32-bit machines. Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> --- include/linux/energy_model.h | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/include/linux/energy_model.h b/include/linux/energy_model.h index 3f221dbf5f95..2016f5a706e0 100644 --- a/include/linux/energy_model.h +++ b/include/linux/energy_model.h @@ -101,7 +101,7 @@ void em_dev_unregister_perf_domain(struct device *dev); * Return: the sum of the energy consumed by the CPUs of the domain assuming * a capacity state satisfying the max utilization of the domain. */ -static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, +static inline u64 em_cpu_energy(struct em_perf_domain *pd, unsigned long max_util, unsigned long sum_util, unsigned long allowed_cpu_cap) { @@ -180,7 +180,7 @@ static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, * pd_nrg = ------------------------ (4) * scale_cpu */ - return ps->cost * sum_util / scale_cpu; + return div_u64((u64)ps->cost * sum_util, scale_cpu); } /** @@ -217,7 +217,7 @@ static inline struct em_perf_domain *em_pd_get(struct device *dev) { return NULL; } -static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, +static inline u64 em_cpu_energy(struct em_perf_domain *pd, unsigned long max_util, unsigned long sum_util, unsigned long allowed_cpu_cap) { -- 2.17.1
The Energy Model (EM) provides useful information about device power in each performance state to other subsystems like: Energy Aware Scheduler (EAS). The energy calculation in EAS does arithmetic operation based on the EM em_cpu_energy(). Current implementation of that function uses em_perf_state::cost as a pre-computed cost coefficient equal to: cost = power * max_frequency / frequency. The 'power' is expressed in milli-Watts (or in abstract scale). There are corner cases then the EAS energy calculation for two Performance Domains (PDs) return the same value, e.g. 10mW. The EAS compares these values to choose smaller one. It might happen that this values are equal due to rounding error. In such scenario, we need better precision, e.g. 10000 times better. To provide this possibility increase the precision on the em_perf_state::cost. This patch allows to avoid the rounding to milli-Watt errors, which might occur in EAS energy estimation for each Performance Domains (PD). The rounding error is common for small tasks which have small utilization values. The rest of the EM code doesn't change, em_perf_state::power is still expressed in milli-Watts (or in abstract scale). Thus, all existing platforms don't have to change their reported power. The same applies to EM clients, like thermal or DTPM (they use em_perf_state::power). Reported-by: CCJ Yeh <CCj.Yeh@mediatek.com> Suggested-by: CCJ Yeh <CCj.Yeh@mediatek.com> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> --- include/linux/energy_model.h | 5 ++++- kernel/power/energy_model.c | 3 ++- 2 files changed, 6 insertions(+), 2 deletions(-) diff --git a/include/linux/energy_model.h b/include/linux/energy_model.h index 2016f5a706e0..91037dd57e61 100644 --- a/include/linux/energy_model.h +++ b/include/linux/energy_model.h @@ -16,7 +16,10 @@ * @power: The power consumed at this level (by 1 CPU or by a registered * device). It can be a total power: static and dynamic. * @cost: The cost coefficient associated with this level, used during - * energy calculation. Equal to: power * max_frequency / frequency + * energy calculation. Equal to: + power * 10000 * max_frequency / frequency + * To increase the energy estimation presision use different + * scale in this coefficient than in @power field. */ struct em_perf_state { unsigned long frequency; diff --git a/kernel/power/energy_model.c b/kernel/power/energy_model.c index 0f4530b3a8cd..2724f0ac417d 100644 --- a/kernel/power/energy_model.c +++ b/kernel/power/energy_model.c @@ -170,7 +170,8 @@ static int em_create_perf_table(struct device *dev, struct em_perf_domain *pd, /* Compute the cost of each performance state. */ fmax = (u64) table[nr_states - 1].frequency; for (i = 0; i < nr_states; i++) { - table[i].cost = div64_u64(fmax * table[i].power, + u64 power_res = (u64)table[i].power * 10000; + table[i].cost = div64_u64(fmax * power_res, table[i].frequency); } -- 2.17.1
On Fri, Jun 25, 2021 at 5:26 PM Lukasz Luba <lukasz.luba@arm.com> wrote: > > The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up > task. It probes many possibilities and compares the estimated energy values > for different scenarios. For calculating those energy values it relies on > Energy Model (EM) data and em_cpu_energy(). The precision which is used in > EM data is in milli-Watts (or abstract scale), which sometimes is not > sufficient. In some cases it might happen that two CPUs from different > Performance Domains (PDs) get the same calculated value for a given task > placement, but in more precised scale, they might differ. This rounding > error has to be addressed. This patch prepares EAS code for better > precision in the coming EM improvements. > > Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> If you want me to pick up this series, this patch requires an ACK from the scheduler maintainers. > --- > kernel/sched/fair.c | 13 +++++++------ > 1 file changed, 7 insertions(+), 6 deletions(-) > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c > index 7b8990fd4896..b517c9e79768 100644 > --- a/kernel/sched/fair.c > +++ b/kernel/sched/fair.c > @@ -6582,7 +6582,7 @@ static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) > * to compute what would be the energy if we decided to actually migrate that > * task. > */ > -static long > +static u64 > compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > { > struct cpumask *pd_mask = perf_domain_span(pd); > @@ -6689,12 +6689,13 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > */ > static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) > { > - unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX; > struct root_domain *rd = cpu_rq(smp_processor_id())->rd; > + u64 prev_delta = ULLONG_MAX, best_delta = ULLONG_MAX; > int cpu, best_energy_cpu = prev_cpu, target = -1; > - unsigned long cpu_cap, util, base_energy = 0; > + unsigned long cpu_cap, util; > struct sched_domain *sd; > struct perf_domain *pd; > + u64 base_energy = 0; > > rcu_read_lock(); > pd = rcu_dereference(rd->pd); > @@ -6718,9 +6719,9 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) > goto unlock; > > for (; pd; pd = pd->next) { > - unsigned long cur_delta, spare_cap, max_spare_cap = 0; > + unsigned long spare_cap, max_spare_cap = 0; > bool compute_prev_delta = false; > - unsigned long base_energy_pd; > + u64 base_energy_pd, cur_delta; > int max_spare_cap_cpu = -1; > > for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) { > @@ -6790,7 +6791,7 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) > * Pick the best CPU if prev_cpu cannot be used, or if it saves at > * least 6% of the energy used by prev_cpu. > */ > - if ((prev_delta == ULONG_MAX) || > + if ((prev_delta == ULLONG_MAX) || > (prev_delta - best_delta) > ((prev_delta + base_energy) >> 4)) > target = best_energy_cpu; > > -- > 2.17.1 >
On 6/30/21 6:01 PM, Rafael J. Wysocki wrote:
> On Fri, Jun 25, 2021 at 5:26 PM Lukasz Luba <lukasz.luba@arm.com> wrote:
>>
>> The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up
>> task. It probes many possibilities and compares the estimated energy values
>> for different scenarios. For calculating those energy values it relies on
>> Energy Model (EM) data and em_cpu_energy(). The precision which is used in
>> EM data is in milli-Watts (or abstract scale), which sometimes is not
>> sufficient. In some cases it might happen that two CPUs from different
>> Performance Domains (PDs) get the same calculated value for a given task
>> placement, but in more precised scale, they might differ. This rounding
>> error has to be addressed. This patch prepares EAS code for better
>> precision in the coming EM improvements.
>>
>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
>
> If you want me to pick up this series, this patch requires an ACK from
> the scheduler maintainers.
>
It would be great, if you could take it after e.g. Peter ACK it.
Peter could you have a look at it, please?
In this patch 1/3 we have only variables upgrade.
Regards,
Lukasz
Hi Peter,
Gentle ping.
You might missed my previous email.
On 6/30/21 6:28 PM, Lukasz Luba wrote:
>
>
> On 6/30/21 6:01 PM, Rafael J. Wysocki wrote:
>> On Fri, Jun 25, 2021 at 5:26 PM Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>
>>> The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up
>>> task. It probes many possibilities and compares the estimated energy
>>> values
>>> for different scenarios. For calculating those energy values it
>>> relies on
>>> Energy Model (EM) data and em_cpu_energy(). The precision which is
>>> used in
>>> EM data is in milli-Watts (or abstract scale), which sometimes is not
>>> sufficient. In some cases it might happen that two CPUs from different
>>> Performance Domains (PDs) get the same calculated value for a given task
>>> placement, but in more precised scale, they might differ. This rounding
>>> error has to be addressed. This patch prepares EAS code for better
>>> precision in the coming EM improvements.
>>>
>>> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
>>
>> If you want me to pick up this series, this patch requires an ACK from
>> the scheduler maintainers.
>>
>
> It would be great, if you could take it after e.g. Peter ACK it.
>
> Peter could you have a look at it, please?
> In this patch 1/3 we have only variables upgrade.
>
> Regards,
> Lukasz
Could you have a look and ACK the patch 1/3, please?
On 25/06/2021 17:26, Lukasz Luba wrote: > The Energy Model (EM) em_cpu_energy() is responsible for providing good > estimation regarding CPUs energy. It contains proper data structures which > are then used during calculation. The values stored in there are in > milli-Watts precision (or in abstract scale) smaller that 0xffff, which use I guess you refer to 'if (... || power > EM_MAX_POWER)' check in em_create_perf_table() [kernel/power/energy_model.c]. > sufficient unsigned long even on 32-bit machines. There are scenarios where ^^^^^^^^^ Can you describe these scenarios better with one example (EAS placement of an example task on a 2 PD system) which highlights the issue and how it this patch-set solves it? In this example you can list all the things which must be there to create a situation in EAS in which the patch-set helps. > we would like to provide calculated estimations in a better precision and > the values might be 1000 times bigger. This patch makes possible to use Where is this `1000` coming from? > quite big values for also 32-bit machines. > > Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> > --- > include/linux/energy_model.h | 6 +++--- > 1 file changed, 3 insertions(+), 3 deletions(-) > > diff --git a/include/linux/energy_model.h b/include/linux/energy_model.h > index 3f221dbf5f95..2016f5a706e0 100644 > --- a/include/linux/energy_model.h > +++ b/include/linux/energy_model.h > @@ -101,7 +101,7 @@ void em_dev_unregister_perf_domain(struct device *dev); > * Return: the sum of the energy consumed by the CPUs of the domain assuming > * a capacity state satisfying the max utilization of the domain. > */ > -static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, > +static inline u64 em_cpu_energy(struct em_perf_domain *pd, > unsigned long max_util, unsigned long sum_util, > unsigned long allowed_cpu_cap) > { > @@ -180,7 +180,7 @@ static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, > * pd_nrg = ------------------------ (4) > * scale_cpu > */ > - return ps->cost * sum_util / scale_cpu; > + return div_u64((u64)ps->cost * sum_util, scale_cpu); > } > > /** > @@ -217,7 +217,7 @@ static inline struct em_perf_domain *em_pd_get(struct device *dev) > { > return NULL; > } > -static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, > +static inline u64 em_cpu_energy(struct em_perf_domain *pd, > unsigned long max_util, unsigned long sum_util, > unsigned long allowed_cpu_cap) > { >
On 25/06/2021 17:26, Lukasz Luba wrote: > The Energy Model (EM) provides useful information about device power in > each performance state to other subsystems like: Energy Aware Scheduler > (EAS). The energy calculation in EAS does arithmetic operation based on > the EM em_cpu_energy(). Current implementation of that function uses > em_perf_state::cost as a pre-computed cost coefficient equal to: > cost = power * max_frequency / frequency. > The 'power' is expressed in milli-Watts (or in abstract scale). > > There are corner cases then the EAS energy calculation for two Performance ^^^^^^^^^^^^ Again, an easy to understand example to describe in which situation this change would bring a benefit would help. > Domains (PDs) return the same value, e.g. 10mW. The EAS compares these > values to choose smaller one. It might happen that this values are equal > due to rounding error. In such scenario, we need better precision, e.g. > 10000 times better. To provide this possibility increase the precision on > the em_perf_state::cost. > > This patch allows to avoid the rounding to milli-Watt errors, which might > occur in EAS energy estimation for each Performance Domains (PD). The > rounding error is common for small tasks which have small utilization > values. What's the influence of the CPU utilization 'cpu_util_next()' here? compute_energy() em_cpu_energy() return ps->cost * sum_util / scale_cpu ^^^^^^^^ > The rest of the EM code doesn't change, em_perf_state::power is still > expressed in milli-Watts (or in abstract scale). Thus, all existing > platforms don't have to change their reported power. The same applies to Not only existing platforms since there are no changes. So why highlighting `existing` here.? > EM clients, like thermal or DTPM (they use em_perf_state::power). > > Reported-by: CCJ Yeh <CCj.Yeh@mediatek.com> > Suggested-by: CCJ Yeh <CCj.Yeh@mediatek.com> > Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> > --- > include/linux/energy_model.h | 5 ++++- > kernel/power/energy_model.c | 3 ++- > 2 files changed, 6 insertions(+), 2 deletions(-) > > diff --git a/include/linux/energy_model.h b/include/linux/energy_model.h > index 2016f5a706e0..91037dd57e61 100644 > --- a/include/linux/energy_model.h > +++ b/include/linux/energy_model.h > @@ -16,7 +16,10 @@ > * @power: The power consumed at this level (by 1 CPU or by a registered > * device). It can be a total power: static and dynamic. > * @cost: The cost coefficient associated with this level, used during > - * energy calculation. Equal to: power * max_frequency / frequency > + * energy calculation. Equal to: > + power * 10000 * max_frequency / frequency > + * To increase the energy estimation presision use different > + * scale in this coefficient than in @power field. > */ > struct em_perf_state { > unsigned long frequency; > diff --git a/kernel/power/energy_model.c b/kernel/power/energy_model.c > index 0f4530b3a8cd..2724f0ac417d 100644 > --- a/kernel/power/energy_model.c > +++ b/kernel/power/energy_model.c > @@ -170,7 +170,8 @@ static int em_create_perf_table(struct device *dev, struct em_perf_domain *pd, > /* Compute the cost of each performance state. */ > fmax = (u64) table[nr_states - 1].frequency; > for (i = 0; i < nr_states; i++) { > - table[i].cost = div64_u64(fmax * table[i].power, > + u64 power_res = (u64)table[i].power * 10000; > + table[i].cost = div64_u64(fmax * power_res, > table[i].frequency); > } > >
On 7/5/21 1:44 PM, Dietmar Eggemann wrote: > On 25/06/2021 17:26, Lukasz Luba wrote: >> The Energy Model (EM) em_cpu_energy() is responsible for providing good >> estimation regarding CPUs energy. It contains proper data structures which >> are then used during calculation. The values stored in there are in >> milli-Watts precision (or in abstract scale) smaller that 0xffff, which use > > I guess you refer to 'if (... || power > EM_MAX_POWER)' check in > em_create_perf_table() [kernel/power/energy_model.c]. Correct > >> sufficient unsigned long even on 32-bit machines. There are scenarios where > ^^^^^^^^^ > > Can you describe these scenarios better with one example (EAS placement > of an example task on a 2 PD system) which highlights the issue and how > it this patch-set solves it? There are two places in the code where it makes a difference: 1. In the find_energy_efficient_cpu() where we are searching for best_delta. We might suffer there when two PDs return the same result, like in the example below. Scenario: Low utilized system e.g. ~200 sum_util for PD0 and ~220 for PD1. There are quite a few small tasks ~10-15 util. These tasks would suffer for the rounding error. Such system utilization has been seen while playing some simple games. In such condition our partner reported 5..10mA less battery drain. Some details: We have two Perf Domains (PDs): PD0 (big) and PD1 (little) Let's compare w/o patch set ('old') and w/ patch set ('new') We are comparing energy w/ task and w/o task placed in the PDs a) 'old' w/o patch set, PD0 task_util = 13 cost = 480 sum_util_w/o_task = 215 sum_util_w_task = 228 scale_cpu = 1024 energy_w/o_task = 480 * 215 / 1024 = 100.78 => 100 energy_w_task = 480 * 228 / 1024 = 106.87 => 106 energy_diff = 106 - 100 = 6 (this is equal to 'old' PD1's energy_diff in 'c)') b) 'new' w/ patch set, PD0 task_util = 13 cost = 480 * 10000 = 4800000 sum_util_w/o_task = 215 sum_util_w_task = 228 energy_w/o_task = 4800000 * 215 / 1024 = 1007812 energy_w_task = 4800000 * 228 / 1024 = 1068750 energy_diff = 1068750 - 1007812 = 60938 (this is not equal to 'new' PD1's energy_diff in 'd)') c) 'old' w/o patch set, PD1 task_util = 13 cost = 160 sum_util_w/o_task = 283 sum_util_w_task = 293 scale_cpu = 355 energy_w/o_task = 160 * 283 / 355 = 127.55 => 127 energy_w_task = 160 * 296 / 355 = 133.41 => 133 energy_diff = 133 - 127 = 6 (this is equal to 'old' PD0's energy_diff in 'a)') d) 'new' w/ patch set, PD1 task_util = 13 cost = 160 * 10000 = 1600000 sum_util_w/o_task = 283 sum_util_w_task = 293 scale_cpu = 355 (no '/ scale_cpu' needed here) energy_w/o_task = 1600000 * 283 / 355 = 1275492 energy_w_task = 1600000 * 296 / 355 = 1334084 energy_diff = 1334084 - 1275492 = 58592 (this is not equal to 'new' PD0's energy_diff in 'b)') 2. Difference in the the last feec() step: margin filter With the patch set the margin comparison also has better resolution, so it's possible to hit better placement thanks to that. Please see the traces below. How to interpret these values: In the first trace below, there is diff=124964 and margin=123381 the EM 'cost' is multiplied by 10000, so we we divide these two, it will be '12 > 12', so it won't be placed into the better PD with lower best delta. In the last 2 examples you would see close values in the prev_delta=49390 best_delta=43945 Without the patch they would be rounded to prev_delta=4 best_delta=4 and the task might be placed wrongly. +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ systemd-logind-440 [000] d..5 82.164218: compute_energy: energy=43945, sum_util=9 cpu=4 systemd-logind-440 [000] d..5 82.164232: compute_energy: energy=766601, sum_util=157 cpu=4 systemd-logind-440 [000] d..5 82.164242: compute_energy: energy=766601, sum_util=157 cpu=4 systemd-logind-440 [000] d..5 82.164253: compute_energy: energy=1207500, sum_util=299 cpu=0 systemd-logind-440 [000] d..5 82.164263: compute_energy: energy=1805192, sum_util=447 cpu=0 systemd-logind-440 [000] d..5 82.164273: select_task_rq_fair: EAS: prev_delta=722656 best_delta=597692 diff=124964 margin=123381 systemd-logind-440 [000] d..5 82.164278: select_task_rq_fair: EAS: hit!!! systemd-logind-440 [000] d.h4 134.954038: compute_energy: energy=366210, sum_util=75 cpu=4 systemd-logind-440 [000] d.h4 134.954067: compute_energy: energy=463867, sum_util=95 cpu=4 systemd-logind-440 [000] d.h4 134.954090: compute_energy: energy=463867, sum_util=95 cpu=4 systemd-logind-440 [000] d.h4 134.954117: compute_energy: energy=257347, sum_util=99 cpu=0 systemd-logind-440 [000] d.h4 134.954137: compute_energy: energy=309336, sum_util=119 cpu=0 systemd-logind-440 [000] d.h4 134.954160: select_task_rq_fair: EAS: prev_delta=97657 best_delta=51989 diff=45668 margin=45075 systemd-logind-440 [000] d.h4 134.954171: select_task_rq_fair: EAS: hit!!! <idle>-0 [001] d.s4 226.019763: compute_energy: energy=0, sum_util=0 cpu=4 <idle>-0 [001] d.s4 226.019790: compute_energy: energy=43945, sum_util=9 cpu=4 <idle>-0 [001] d.s4 226.019817: compute_energy: energy=5198, sum_util=2 cpu=0 <idle>-0 [001] d.s4 226.019838: compute_energy: energy=54588, sum_util=21 cpu=0 <idle>-0 [001] d.s4 226.019858: compute_energy: energy=54588, sum_util=21 cpu=0 <idle>-0 [001] d.s4 226.019881: select_task_rq_fair: EAS: prev_delta=49390 best_delta=43945 diff=5445 margin=3411 <idle>-0 [001] d.s4 226.019891: select_task_rq_fair: EAS: hit!!! <idle>-0 [001] d.s4 270.019780: compute_energy: energy=0, sum_util=0 cpu=4 <idle>-0 [001] d.s4 270.019807: compute_energy: energy=43945, sum_util=9 cpu=4 <idle>-0 [001] d.s4 270.019833: compute_energy: energy=5198, sum_util=2 cpu=0 <idle>-0 [001] d.s4 270.019854: compute_energy: energy=54588, sum_util=21 cpu=0 <idle>-0 [001] d.s4 270.019874: compute_energy: energy=54588, sum_util=21 cpu=0 <idle>-0 [001] d.s4 270.019897: select_task_rq_fair: EAS: prev_delta=49390 best_delta=43945 diff=5445 margin=3411 <idle>-0 [001] d.s4 270.019908: select_task_rq_fair: EAS: hit!!! +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ > > In this example you can list all the things which must be there to > create a situation in EAS in which the patch-set helps. I hope the description above now add more light into this issue. > >> we would like to provide calculated estimations in a better precision and >> the values might be 1000 times bigger. This patch makes possible to use > > Where is this `1000` coming from? It's just a statement that in the next patches we would increase the resolution by a few orders of magnitude. In patch 3/3 it's 10000. I can align with that value also in this statement. Thank you Dietmar for having a look at this! Regards, Lukasz
On 7/5/21 1:45 PM, Dietmar Eggemann wrote: > On 25/06/2021 17:26, Lukasz Luba wrote: >> The Energy Model (EM) provides useful information about device power in >> each performance state to other subsystems like: Energy Aware Scheduler >> (EAS). The energy calculation in EAS does arithmetic operation based on >> the EM em_cpu_energy(). Current implementation of that function uses >> em_perf_state::cost as a pre-computed cost coefficient equal to: >> cost = power * max_frequency / frequency. >> The 'power' is expressed in milli-Watts (or in abstract scale). >> >> There are corner cases then the EAS energy calculation for two Performance > ^^^^^^^^^^^^ > > Again, an easy to understand example to describe in which situation this > change would bring a benefit would help. > >> Domains (PDs) return the same value, e.g. 10mW. The EAS compares these >> values to choose smaller one. It might happen that this values are equal >> due to rounding error. In such scenario, we need better precision, e.g. >> 10000 times better. To provide this possibility increase the precision on >> the em_perf_state::cost. >> >> This patch allows to avoid the rounding to milli-Watt errors, which might >> occur in EAS energy estimation for each Performance Domains (PD). The >> rounding error is common for small tasks which have small utilization >> values. > > What's the influence of the CPU utilization 'cpu_util_next()' here? > > compute_energy() > em_cpu_energy() > return ps->cost * sum_util / scale_cpu > ^^^^^^^^ This is the place where the rounding error triggers. If sum_util is small and scale_cpu is e.g. 1024, then we have a small fraction here. It depends on the EM 'cost', but for most platforms we have small power and cost values, so we suffer this rounding. The example that I gave in my response in patch 2/3 shows this. >> The rest of the EM code doesn't change, em_perf_state::power is still >> expressed in milli-Watts (or in abstract scale). Thus, all existing >> platforms don't have to change their reported power. The same applies to > > Not only existing platforms since there are no changes. So why > highlighting `existing` here.? I just wanted to be clear that it doesn't affect existing platforms at all. We don't require to report power in better resolution e.g. micro-Watts. Also, the clients in the kernel won't be affected, since they use EM 'power' filed, not 'cost'.
On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@arm.com> wrote: > > The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up > task. It probes many possibilities and compares the estimated energy values > for different scenarios. For calculating those energy values it relies on > Energy Model (EM) data and em_cpu_energy(). The precision which is used in > EM data is in milli-Watts (or abstract scale), which sometimes is not > sufficient. In some cases it might happen that two CPUs from different > Performance Domains (PDs) get the same calculated value for a given task > placement, but in more precised scale, they might differ. This rounding > error has to be addressed. This patch prepares EAS code for better > precision in the coming EM improvements. Could you explain why 32bits results are not enough and you need to move to 64bits ? Right now the result is in the range [0..2^32[ mW. If you need more precision and you want to return uW instead, you will have a result in the range [0..4kW[ which seems to be still enough > > Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> > --- > kernel/sched/fair.c | 13 +++++++------ > 1 file changed, 7 insertions(+), 6 deletions(-) > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c > index 7b8990fd4896..b517c9e79768 100644 > --- a/kernel/sched/fair.c > +++ b/kernel/sched/fair.c > @@ -6582,7 +6582,7 @@ static unsigned long cpu_util_next(int cpu, struct task_struct *p, int dst_cpu) > * to compute what would be the energy if we decided to actually migrate that > * task. > */ > -static long > +static u64 > compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > { > struct cpumask *pd_mask = perf_domain_span(pd); > @@ -6689,12 +6689,13 @@ compute_energy(struct task_struct *p, int dst_cpu, struct perf_domain *pd) > */ > static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) > { > - unsigned long prev_delta = ULONG_MAX, best_delta = ULONG_MAX; > struct root_domain *rd = cpu_rq(smp_processor_id())->rd; > + u64 prev_delta = ULLONG_MAX, best_delta = ULLONG_MAX; > int cpu, best_energy_cpu = prev_cpu, target = -1; > - unsigned long cpu_cap, util, base_energy = 0; > + unsigned long cpu_cap, util; > struct sched_domain *sd; > struct perf_domain *pd; > + u64 base_energy = 0; > > rcu_read_lock(); > pd = rcu_dereference(rd->pd); > @@ -6718,9 +6719,9 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) > goto unlock; > > for (; pd; pd = pd->next) { > - unsigned long cur_delta, spare_cap, max_spare_cap = 0; > + unsigned long spare_cap, max_spare_cap = 0; > bool compute_prev_delta = false; > - unsigned long base_energy_pd; > + u64 base_energy_pd, cur_delta; > int max_spare_cap_cpu = -1; > > for_each_cpu_and(cpu, perf_domain_span(pd), sched_domain_span(sd)) { > @@ -6790,7 +6791,7 @@ static int find_energy_efficient_cpu(struct task_struct *p, int prev_cpu) > * Pick the best CPU if prev_cpu cannot be used, or if it saves at > * least 6% of the energy used by prev_cpu. > */ > - if ((prev_delta == ULONG_MAX) || > + if ((prev_delta == ULLONG_MAX) || > (prev_delta - best_delta) > ((prev_delta + base_energy) >> 4)) > target = best_energy_cpu; > > -- > 2.17.1 >
On Fri, Jun 25, 2021 at 04:26:02PM +0100, Lukasz Luba wrote:
> The Energy Model (EM) em_cpu_energy() is responsible for providing good
> estimation regarding CPUs energy. It contains proper data structures which
> are then used during calculation. The values stored in there are in
> milli-Watts precision (or in abstract scale) smaller that 0xffff, which use
> sufficient unsigned long even on 32-bit machines. There are scenarios where
> we would like to provide calculated estimations in a better precision and
> the values might be 1000 times bigger. This patch makes possible to use
> quite big values for also 32-bit machines.
>
> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com>
> ---
> include/linux/energy_model.h | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/include/linux/energy_model.h b/include/linux/energy_model.h
> index 3f221dbf5f95..2016f5a706e0 100644
> --- a/include/linux/energy_model.h
> +++ b/include/linux/energy_model.h
> @@ -101,7 +101,7 @@ void em_dev_unregister_perf_domain(struct device *dev);
> * Return: the sum of the energy consumed by the CPUs of the domain assuming
> * a capacity state satisfying the max utilization of the domain.
> */
> -static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
> +static inline u64 em_cpu_energy(struct em_perf_domain *pd,
> unsigned long max_util, unsigned long sum_util,
> unsigned long allowed_cpu_cap)
> {
> @@ -180,7 +180,7 @@ static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
> * pd_nrg = ------------------------ (4)
> * scale_cpu
> */
> - return ps->cost * sum_util / scale_cpu;
> + return div_u64((u64)ps->cost * sum_util, scale_cpu);
So these patches are all rather straight forward, however.. the above is
pretty horrific on a 32bit box, and we do quite a few of them per
wakeup. Is this really worth the performance penalty on 32bit CPUs?
Do you really still care about 32bit CPUs, or is this mostly an artifact
of wanting to unconditionally increase the precision?
On 7/7/21 8:07 AM, Vincent Guittot wrote:
> On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>
>> The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up
>> task. It probes many possibilities and compares the estimated energy values
>> for different scenarios. For calculating those energy values it relies on
>> Energy Model (EM) data and em_cpu_energy(). The precision which is used in
>> EM data is in milli-Watts (or abstract scale), which sometimes is not
>> sufficient. In some cases it might happen that two CPUs from different
>> Performance Domains (PDs) get the same calculated value for a given task
>> placement, but in more precised scale, they might differ. This rounding
>> error has to be addressed. This patch prepares EAS code for better
>> precision in the coming EM improvements.
>
> Could you explain why 32bits results are not enough and you need to
> move to 64bits ?
>
> Right now the result is in the range [0..2^32[ mW. If you need more
> precision and you want to return uW instead, you will have a result in
> the range [0..4kW[ which seems to be still enough
>
Currently we have the max value limit for 'power' in EM which is
EM_MAX_POWER 0xffff (64k - 1). We allow to register such big power
values ~64k mW (~64Watts) for an OPP. Then based on 'power' we
pre-calculate 'cost' fields:
cost[i] = power[i] * freq_max / freq[i]
So, for max freq the cost == power. Let's use that in the example.
Then the em_cpu_energy() calculates as follow:
cost * sum_util / scale_cpu
We are interested in the first part - the value of multiplication.
The sum_util values that we can see for x CPUs which have scale_cap=1024
can be close to 800, let's use it in the example:
cost * sum_util = 64k * (x * 800), where
x=4: ~200mln
x=8: ~400mln
x=16: ~800mln
x=64: ~3200mln (last one which would fit in u32)
When we increase the precision by even 100, then the above values won't
fit in the u32. Even a max cost of e.g. 10k mW and 100 precision has
issues:
cost * sum_util = (10k *100) * (x * 800), where
x=4: ~3200mln
x=8: ~6400mln
For *1000 precision even a power of 1Watt becomes an issue:
cost * sum_util = (1k *1000) * (x * 800), where
x=4: ~3200mln
x=8: ~6400mln
That's why to make the code safe for bigger power values, I had to use
the u64 on 32bit machines.
On Wed, 7 Jul 2021 at 09:49, Lukasz Luba <lukasz.luba@arm.com> wrote: > > > > On 7/7/21 8:07 AM, Vincent Guittot wrote: > > On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@arm.com> wrote: > >> > >> The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up > >> task. It probes many possibilities and compares the estimated energy values > >> for different scenarios. For calculating those energy values it relies on > >> Energy Model (EM) data and em_cpu_energy(). The precision which is used in > >> EM data is in milli-Watts (or abstract scale), which sometimes is not > >> sufficient. In some cases it might happen that two CPUs from different > >> Performance Domains (PDs) get the same calculated value for a given task > >> placement, but in more precised scale, they might differ. This rounding > >> error has to be addressed. This patch prepares EAS code for better > >> precision in the coming EM improvements. > > > > Could you explain why 32bits results are not enough and you need to > > move to 64bits ? > > > > Right now the result is in the range [0..2^32[ mW. If you need more > > precision and you want to return uW instead, you will have a result in > > the range [0..4kW[ which seems to be still enough > > > > Currently we have the max value limit for 'power' in EM which is > EM_MAX_POWER 0xffff (64k - 1). We allow to register such big power > values ~64k mW (~64Watts) for an OPP. Then based on 'power' we > pre-calculate 'cost' fields: > cost[i] = power[i] * freq_max / freq[i] > So, for max freq the cost == power. Let's use that in the example. > > Then the em_cpu_energy() calculates as follow: > cost * sum_util / scale_cpu > We are interested in the first part - the value of multiplication. But all these are internal computations of the energy model. At the end, the computed energy that is returned by compute_energy() and em_cpu_energy(), fits in a long > > The sum_util values that we can see for x CPUs which have scale_cap=1024 > can be close to 800, let's use it in the example: > cost * sum_util = 64k * (x * 800), where > x=4: ~200mln > x=8: ~400mln > x=16: ~800mln > x=64: ~3200mln (last one which would fit in u32) > > When we increase the precision by even 100, then the above values won't > fit in the u32. Even a max cost of e.g. 10k mW and 100 precision has > issues: > cost * sum_util = (10k *100) * (x * 800), where > x=4: ~3200mln > x=8: ~6400mln > > For *1000 precision even a power of 1Watt becomes an issue: > cost * sum_util = (1k *1000) * (x * 800), where > x=4: ~3200mln > x=8: ~6400mln > > That's why to make the code safe for bigger power values, I had to use > the u64 on 32bit machines.
On 7/7/21 8:07 AM, Peter Zijlstra wrote: > On Fri, Jun 25, 2021 at 04:26:02PM +0100, Lukasz Luba wrote: >> The Energy Model (EM) em_cpu_energy() is responsible for providing good >> estimation regarding CPUs energy. It contains proper data structures which >> are then used during calculation. The values stored in there are in >> milli-Watts precision (or in abstract scale) smaller that 0xffff, which use >> sufficient unsigned long even on 32-bit machines. There are scenarios where >> we would like to provide calculated estimations in a better precision and >> the values might be 1000 times bigger. This patch makes possible to use >> quite big values for also 32-bit machines. >> >> Signed-off-by: Lukasz Luba <lukasz.luba@arm.com> >> --- >> include/linux/energy_model.h | 6 +++--- >> 1 file changed, 3 insertions(+), 3 deletions(-) >> >> diff --git a/include/linux/energy_model.h b/include/linux/energy_model.h >> index 3f221dbf5f95..2016f5a706e0 100644 >> --- a/include/linux/energy_model.h >> +++ b/include/linux/energy_model.h >> @@ -101,7 +101,7 @@ void em_dev_unregister_perf_domain(struct device *dev); >> * Return: the sum of the energy consumed by the CPUs of the domain assuming >> * a capacity state satisfying the max utilization of the domain. >> */ >> -static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, >> +static inline u64 em_cpu_energy(struct em_perf_domain *pd, >> unsigned long max_util, unsigned long sum_util, >> unsigned long allowed_cpu_cap) >> { >> @@ -180,7 +180,7 @@ static inline unsigned long em_cpu_energy(struct em_perf_domain *pd, >> * pd_nrg = ------------------------ (4) >> * scale_cpu >> */ >> - return ps->cost * sum_util / scale_cpu; >> + return div_u64((u64)ps->cost * sum_util, scale_cpu); > > So these patches are all rather straight forward, however.. the above is > pretty horrific on a 32bit box, and we do quite a few of them per > wakeup. Is this really worth the performance penalty on 32bit CPUs? True, for 2 cluster SoC we might do this 5 times (or less, depends on system state). We don't have new 32bit big.LITTLE platforms, the newest is ~7years old and is actually the only one using EAS. It's not put into new devices AFAIK. > > Do you really still care about 32bit CPUs, or is this mostly an artifact > of wanting to unconditionally increase the precision? > We discussed this internally and weighted the 32bit old big.little. There is a solution, but needs more work and a lot of changes in the whole kernel due to modified EM (affects IPA, DTPM, registration, ...). I have been working on a next step for code that you've pointed: get rid of this runtime division. It would be possible to pre-calculate the: 'ps->cost / scale_cpu' at the moment when EM is registered and store it in the ps->cost. So we would have just: return ps->cost * sum_util The only issue is a late boot of biggest cores, which would destroy the old scale_cpu values for other PDs. I need to probably add RCU locking into the EM and update the other PDs' EMs when the last biggest CPU boots after a few second and registers its EM. For now we would live with this simple code which improves all recent 64bit platforms and is easy to take it into Android common kernel. The next step would be more scattered across other subsystems, so harder to backport to Android 5.4 and others.
On 7/7/21 9:00 AM, Vincent Guittot wrote: > On Wed, 7 Jul 2021 at 09:49, Lukasz Luba <lukasz.luba@arm.com> wrote: >> >> >> >> On 7/7/21 8:07 AM, Vincent Guittot wrote: >>> On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@arm.com> wrote: >>>> >>>> The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up >>>> task. It probes many possibilities and compares the estimated energy values >>>> for different scenarios. For calculating those energy values it relies on >>>> Energy Model (EM) data and em_cpu_energy(). The precision which is used in >>>> EM data is in milli-Watts (or abstract scale), which sometimes is not >>>> sufficient. In some cases it might happen that two CPUs from different >>>> Performance Domains (PDs) get the same calculated value for a given task >>>> placement, but in more precised scale, they might differ. This rounding >>>> error has to be addressed. This patch prepares EAS code for better >>>> precision in the coming EM improvements. >>> >>> Could you explain why 32bits results are not enough and you need to >>> move to 64bits ? >>> >>> Right now the result is in the range [0..2^32[ mW. If you need more >>> precision and you want to return uW instead, you will have a result in >>> the range [0..4kW[ which seems to be still enough >>> >> >> Currently we have the max value limit for 'power' in EM which is >> EM_MAX_POWER 0xffff (64k - 1). We allow to register such big power >> values ~64k mW (~64Watts) for an OPP. Then based on 'power' we >> pre-calculate 'cost' fields: >> cost[i] = power[i] * freq_max / freq[i] >> So, for max freq the cost == power. Let's use that in the example. >> >> Then the em_cpu_energy() calculates as follow: >> cost * sum_util / scale_cpu >> We are interested in the first part - the value of multiplication. > > But all these are internal computations of the energy model. At the > end, the computed energy that is returned by compute_energy() and > em_cpu_energy(), fits in a long Let's take a look at existing *10000 precision for x CPUs: cost * sum_util / scale_cpu = (64k *10000) * (x * 800) / 1024 which is: x * ~500mln So to be close to overflowing u32 the 'x' has to be > (?=) 8 (depends on sum_util). > >> >> The sum_util values that we can see for x CPUs which have scale_cap=1024 >> can be close to 800, let's use it in the example: >> cost * sum_util = 64k * (x * 800), where >> x=4: ~200mln >> x=8: ~400mln >> x=16: ~800mln >> x=64: ~3200mln (last one which would fit in u32) >> >> When we increase the precision by even 100, then the above values won't >> fit in the u32. Even a max cost of e.g. 10k mW and 100 precision has >> issues: >> cost * sum_util = (10k *100) * (x * 800), where >> x=4: ~3200mln >> x=8: ~6400mln >> >> For *1000 precision even a power of 1Watt becomes an issue: >> cost * sum_util = (1k *1000) * (x * 800), where >> x=4: ~3200mln >> x=8: ~6400mln >> >> That's why to make the code safe for bigger power values, I had to use >> the u64 on 32bit machines.
On Wed, 7 Jul 2021 at 10:23, Lukasz Luba <lukasz.luba@arm.com> wrote: > > > > On 7/7/21 9:00 AM, Vincent Guittot wrote: > > On Wed, 7 Jul 2021 at 09:49, Lukasz Luba <lukasz.luba@arm.com> wrote: > >> > >> > >> > >> On 7/7/21 8:07 AM, Vincent Guittot wrote: > >>> On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@arm.com> wrote: > >>>> > >>>> The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up > >>>> task. It probes many possibilities and compares the estimated energy values > >>>> for different scenarios. For calculating those energy values it relies on > >>>> Energy Model (EM) data and em_cpu_energy(). The precision which is used in > >>>> EM data is in milli-Watts (or abstract scale), which sometimes is not > >>>> sufficient. In some cases it might happen that two CPUs from different > >>>> Performance Domains (PDs) get the same calculated value for a given task > >>>> placement, but in more precised scale, they might differ. This rounding > >>>> error has to be addressed. This patch prepares EAS code for better > >>>> precision in the coming EM improvements. > >>> > >>> Could you explain why 32bits results are not enough and you need to > >>> move to 64bits ? > >>> > >>> Right now the result is in the range [0..2^32[ mW. If you need more > >>> precision and you want to return uW instead, you will have a result in > >>> the range [0..4kW[ which seems to be still enough > >>> > >> > >> Currently we have the max value limit for 'power' in EM which is > >> EM_MAX_POWER 0xffff (64k - 1). We allow to register such big power > >> values ~64k mW (~64Watts) for an OPP. Then based on 'power' we > >> pre-calculate 'cost' fields: > >> cost[i] = power[i] * freq_max / freq[i] > >> So, for max freq the cost == power. Let's use that in the example. > >> > >> Then the em_cpu_energy() calculates as follow: > >> cost * sum_util / scale_cpu > >> We are interested in the first part - the value of multiplication. > > > > But all these are internal computations of the energy model. At the > > end, the computed energy that is returned by compute_energy() and > > em_cpu_energy(), fits in a long > > Let's take a look at existing *10000 precision for x CPUs: > cost * sum_util / scale_cpu = > (64k *10000) * (x * 800) / 1024 > which is: > x * ~500mln > > So to be close to overflowing u32 the 'x' has to be > (?=) 8 > (depends on sum_util). Sorry but I don't get your point. This patch is about the return type of compute_energy() and em_cpu_energy(). And even if we decide to return uW instead of mW, there is still a lot of margin. It's not because you need u64 for computing intermediate value that you must returns u64 > > > > >> > >> The sum_util values that we can see for x CPUs which have scale_cap=1024 > >> can be close to 800, let's use it in the example: > >> cost * sum_util = 64k * (x * 800), where > >> x=4: ~200mln > >> x=8: ~400mln > >> x=16: ~800mln > >> x=64: ~3200mln (last one which would fit in u32) > >> > >> When we increase the precision by even 100, then the above values won't > >> fit in the u32. Even a max cost of e.g. 10k mW and 100 precision has > >> issues: > >> cost * sum_util = (10k *100) * (x * 800), where > >> x=4: ~3200mln > >> x=8: ~6400mln > >> > >> For *1000 precision even a power of 1Watt becomes an issue: > >> cost * sum_util = (1k *1000) * (x * 800), where > >> x=4: ~3200mln > >> x=8: ~6400mln > >> > >> That's why to make the code safe for bigger power values, I had to use > >> the u64 on 32bit machines.
On 07/07/2021 10:23, Lukasz Luba wrote: > > On 7/7/21 9:00 AM, Vincent Guittot wrote: >> On Wed, 7 Jul 2021 at 09:49, Lukasz Luba <lukasz.luba@arm.com> wrote: >>> >>> >>> >>> On 7/7/21 8:07 AM, Vincent Guittot wrote: >>>> On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@arm.com> wrote: [...] >>>> Could you explain why 32bits results are not enough and you need to >>>> move to 64bits ? >>>> >>>> Right now the result is in the range [0..2^32[ mW. If you need more >>>> precision and you want to return uW instead, you will have a result in >>>> the range [0..4kW[ which seems to be still enough >>>> >>> >>> Currently we have the max value limit for 'power' in EM which is >>> EM_MAX_POWER 0xffff (64k - 1). We allow to register such big power >>> values ~64k mW (~64Watts) for an OPP. Then based on 'power' we >>> pre-calculate 'cost' fields: >>> cost[i] = power[i] * freq_max / freq[i] >>> So, for max freq the cost == power. Let's use that in the example. >>> >>> Then the em_cpu_energy() calculates as follow: >>> cost * sum_util / scale_cpu >>> We are interested in the first part - the value of multiplication. >> >> But all these are internal computations of the energy model. At the >> end, the computed energy that is returned by compute_energy() and >> em_cpu_energy(), fits in a long > > Let's take a look at existing *10000 precision for x CPUs: > cost * sum_util / scale_cpu = > (64k *10000) * (x * 800) / 1024 > which is: > x * ~500mln > > So to be close to overflowing u32 the 'x' has to be > (?=) 8 > (depends on sum_util). I assume the worst case is `x * 1024` (max return value of effective_cpu_util = effective_cpu_util()) so x ~ 6.7. I'm not aware of any arm32 b.L. systems with > 4 CPUs in a PD.
On 7/7/21 10:37 AM, Vincent Guittot wrote:
> On Wed, 7 Jul 2021 at 10:23, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>
>>
>>
>> On 7/7/21 9:00 AM, Vincent Guittot wrote:
>>> On Wed, 7 Jul 2021 at 09:49, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>
>>>>
>>>>
>>>> On 7/7/21 8:07 AM, Vincent Guittot wrote:
>>>>> On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>>>
>>>>>> The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up
>>>>>> task. It probes many possibilities and compares the estimated energy values
>>>>>> for different scenarios. For calculating those energy values it relies on
>>>>>> Energy Model (EM) data and em_cpu_energy(). The precision which is used in
>>>>>> EM data is in milli-Watts (or abstract scale), which sometimes is not
>>>>>> sufficient. In some cases it might happen that two CPUs from different
>>>>>> Performance Domains (PDs) get the same calculated value for a given task
>>>>>> placement, but in more precised scale, they might differ. This rounding
>>>>>> error has to be addressed. This patch prepares EAS code for better
>>>>>> precision in the coming EM improvements.
>>>>>
>>>>> Could you explain why 32bits results are not enough and you need to
>>>>> move to 64bits ?
>>>>>
>>>>> Right now the result is in the range [0..2^32[ mW. If you need more
>>>>> precision and you want to return uW instead, you will have a result in
>>>>> the range [0..4kW[ which seems to be still enough
>>>>>
>>>>
>>>> Currently we have the max value limit for 'power' in EM which is
>>>> EM_MAX_POWER 0xffff (64k - 1). We allow to register such big power
>>>> values ~64k mW (~64Watts) for an OPP. Then based on 'power' we
>>>> pre-calculate 'cost' fields:
>>>> cost[i] = power[i] * freq_max / freq[i]
>>>> So, for max freq the cost == power. Let's use that in the example.
>>>>
>>>> Then the em_cpu_energy() calculates as follow:
>>>> cost * sum_util / scale_cpu
>>>> We are interested in the first part - the value of multiplication.
>>>
>>> But all these are internal computations of the energy model. At the
>>> end, the computed energy that is returned by compute_energy() and
>>> em_cpu_energy(), fits in a long
>>
>> Let's take a look at existing *10000 precision for x CPUs:
>> cost * sum_util / scale_cpu =
>> (64k *10000) * (x * 800) / 1024
>> which is:
>> x * ~500mln
>>
>> So to be close to overflowing u32 the 'x' has to be > (?=) 8
>> (depends on sum_util).
>
> Sorry but I don't get your point.
> This patch is about the return type of compute_energy() and
> em_cpu_energy(). And even if we decide to return uW instead of mW,
> there is still a lot of margin.
>
> It's not because you need u64 for computing intermediate value that
> you must returns u64
The example above shows the need of u64 return value for platforms
which are:
- 32bit
- have e.g. 16 CPUs
- has big power value e.g. ~64k mW
Then let's to the calc:
(64k * 10000) * (16 * 800) / 1024 = ~8000mln = ~8bln
The returned value after applying the whole patch set
won't fit in u32 for such cluster.
We might make *assumption* that the 32bit platforms will not
have bigger number of CPUs in the cluster or won't report
big power values. But I didn't wanted to make such assumption.
On 7/7/21 10:45 AM, Dietmar Eggemann wrote:
> On 07/07/2021 10:23, Lukasz Luba wrote:
>>
>> On 7/7/21 9:00 AM, Vincent Guittot wrote:
>>> On Wed, 7 Jul 2021 at 09:49, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>
>>>>
>>>>
>>>> On 7/7/21 8:07 AM, Vincent Guittot wrote:
>>>>> On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@arm.com> wrote:
>
> [...]
>
>>>>> Could you explain why 32bits results are not enough and you need to
>>>>> move to 64bits ?
>>>>>
>>>>> Right now the result is in the range [0..2^32[ mW. If you need more
>>>>> precision and you want to return uW instead, you will have a result in
>>>>> the range [0..4kW[ which seems to be still enough
>>>>>
>>>>
>>>> Currently we have the max value limit for 'power' in EM which is
>>>> EM_MAX_POWER 0xffff (64k - 1). We allow to register such big power
>>>> values ~64k mW (~64Watts) for an OPP. Then based on 'power' we
>>>> pre-calculate 'cost' fields:
>>>> cost[i] = power[i] * freq_max / freq[i]
>>>> So, for max freq the cost == power. Let's use that in the example.
>>>>
>>>> Then the em_cpu_energy() calculates as follow:
>>>> cost * sum_util / scale_cpu
>>>> We are interested in the first part - the value of multiplication.
>>>
>>> But all these are internal computations of the energy model. At the
>>> end, the computed energy that is returned by compute_energy() and
>>> em_cpu_energy(), fits in a long
>>
>> Let's take a look at existing *10000 precision for x CPUs:
>> cost * sum_util / scale_cpu =
>> (64k *10000) * (x * 800) / 1024
>> which is:
>> x * ~500mln
>>
>> So to be close to overflowing u32 the 'x' has to be > (?=) 8
>> (depends on sum_util).
>
> I assume the worst case is `x * 1024` (max return value of
> effective_cpu_util = effective_cpu_util()) so x ~ 6.7.
>
> I'm not aware of any arm32 b.L. systems with > 4 CPUs in a PD.
>
True, arm32 didn't support bigger number than 4 CPUs in the cluster.
We would be safe for them, but I don't want to break with this
assumption any other 32bit platform from competitors, which might
create such 32bit 16cores clusters.
If Peter, Vincent and you are OK to put this assumption about
max safe CPUs number, then we can get rid of patch 1/3.
But the temporary division of u64 must stay, because there is
arm32 platform which need it. So returning also u64 is not a big
harm and looks more consistent.
On Wed, 7 Jul 2021 at 11:48, Lukasz Luba <lukasz.luba@arm.com> wrote: > > > > On 7/7/21 10:37 AM, Vincent Guittot wrote: > > On Wed, 7 Jul 2021 at 10:23, Lukasz Luba <lukasz.luba@arm.com> wrote: > >> > >> > >> > >> On 7/7/21 9:00 AM, Vincent Guittot wrote: > >>> On Wed, 7 Jul 2021 at 09:49, Lukasz Luba <lukasz.luba@arm.com> wrote: > >>>> > >>>> > >>>> > >>>> On 7/7/21 8:07 AM, Vincent Guittot wrote: > >>>>> On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@arm.com> wrote: > >>>>>> > >>>>>> The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up > >>>>>> task. It probes many possibilities and compares the estimated energy values > >>>>>> for different scenarios. For calculating those energy values it relies on > >>>>>> Energy Model (EM) data and em_cpu_energy(). The precision which is used in > >>>>>> EM data is in milli-Watts (or abstract scale), which sometimes is not > >>>>>> sufficient. In some cases it might happen that two CPUs from different > >>>>>> Performance Domains (PDs) get the same calculated value for a given task > >>>>>> placement, but in more precised scale, they might differ. This rounding > >>>>>> error has to be addressed. This patch prepares EAS code for better > >>>>>> precision in the coming EM improvements. > >>>>> > >>>>> Could you explain why 32bits results are not enough and you need to > >>>>> move to 64bits ? > >>>>> > >>>>> Right now the result is in the range [0..2^32[ mW. If you need more > >>>>> precision and you want to return uW instead, you will have a result in > >>>>> the range [0..4kW[ which seems to be still enough > >>>>> > >>>> > >>>> Currently we have the max value limit for 'power' in EM which is > >>>> EM_MAX_POWER 0xffff (64k - 1). We allow to register such big power > >>>> values ~64k mW (~64Watts) for an OPP. Then based on 'power' we > >>>> pre-calculate 'cost' fields: > >>>> cost[i] = power[i] * freq_max / freq[i] > >>>> So, for max freq the cost == power. Let's use that in the example. > >>>> > >>>> Then the em_cpu_energy() calculates as follow: > >>>> cost * sum_util / scale_cpu > >>>> We are interested in the first part - the value of multiplication. > >>> > >>> But all these are internal computations of the energy model. At the > >>> end, the computed energy that is returned by compute_energy() and > >>> em_cpu_energy(), fits in a long > >> > >> Let's take a look at existing *10000 precision for x CPUs: > >> cost * sum_util / scale_cpu = > >> (64k *10000) * (x * 800) / 1024 > >> which is: > >> x * ~500mln > >> > >> So to be close to overflowing u32 the 'x' has to be > (?=) 8 > >> (depends on sum_util). > > > > Sorry but I don't get your point. > > This patch is about the return type of compute_energy() and > > em_cpu_energy(). And even if we decide to return uW instead of mW, > > there is still a lot of margin. > > > > It's not because you need u64 for computing intermediate value that > > you must returns u64 > > The example above shows the need of u64 return value for platforms > which are: > - 32bit > - have e.g. 16 CPUs > - has big power value e.g. ~64k mW > Then let's to the calc: > (64k * 10000) * (16 * 800) / 1024 = ~8000mln = ~8bln so you return a power consumption of 8kW !!! > > The returned value after applying the whole patch set > won't fit in u32 for such cluster. > > We might make *assumption* that the 32bit platforms will not > have bigger number of CPUs in the cluster or won't report > big power values. But I didn't wanted to make such assumption. > >
On Wed, Jul 07, 2021 at 09:09:08AM +0100, Lukasz Luba wrote:
> For now we would live with this simple code which improves
> all recent 64bit platforms and is easy to take it into Android
> common kernel. The next step would be more scattered across
> other subsystems, so harder to backport to Android 5.4 and others.
Ah, you *do* only care about 64bit :-) So one option is to only increase
precision for 64BIT builds, just like we do for scale_load() and
friends.
On 7/7/21 10:56 AM, Vincent Guittot wrote:
> On Wed, 7 Jul 2021 at 11:48, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>
>>
>>
>> On 7/7/21 10:37 AM, Vincent Guittot wrote:
>>> On Wed, 7 Jul 2021 at 10:23, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>
>>>>
>>>>
>>>> On 7/7/21 9:00 AM, Vincent Guittot wrote:
>>>>> On Wed, 7 Jul 2021 at 09:49, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 7/7/21 8:07 AM, Vincent Guittot wrote:
>>>>>>> On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>>>>>
>>>>>>>> The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up
>>>>>>>> task. It probes many possibilities and compares the estimated energy values
>>>>>>>> for different scenarios. For calculating those energy values it relies on
>>>>>>>> Energy Model (EM) data and em_cpu_energy(). The precision which is used in
>>>>>>>> EM data is in milli-Watts (or abstract scale), which sometimes is not
>>>>>>>> sufficient. In some cases it might happen that two CPUs from different
>>>>>>>> Performance Domains (PDs) get the same calculated value for a given task
>>>>>>>> placement, but in more precised scale, they might differ. This rounding
>>>>>>>> error has to be addressed. This patch prepares EAS code for better
>>>>>>>> precision in the coming EM improvements.
>>>>>>>
>>>>>>> Could you explain why 32bits results are not enough and you need to
>>>>>>> move to 64bits ?
>>>>>>>
>>>>>>> Right now the result is in the range [0..2^32[ mW. If you need more
>>>>>>> precision and you want to return uW instead, you will have a result in
>>>>>>> the range [0..4kW[ which seems to be still enough
>>>>>>>
>>>>>>
>>>>>> Currently we have the max value limit for 'power' in EM which is
>>>>>> EM_MAX_POWER 0xffff (64k - 1). We allow to register such big power
>>>>>> values ~64k mW (~64Watts) for an OPP. Then based on 'power' we
>>>>>> pre-calculate 'cost' fields:
>>>>>> cost[i] = power[i] * freq_max / freq[i]
>>>>>> So, for max freq the cost == power. Let's use that in the example.
>>>>>>
>>>>>> Then the em_cpu_energy() calculates as follow:
>>>>>> cost * sum_util / scale_cpu
>>>>>> We are interested in the first part - the value of multiplication.
>>>>>
>>>>> But all these are internal computations of the energy model. At the
>>>>> end, the computed energy that is returned by compute_energy() and
>>>>> em_cpu_energy(), fits in a long
>>>>
>>>> Let's take a look at existing *10000 precision for x CPUs:
>>>> cost * sum_util / scale_cpu =
>>>> (64k *10000) * (x * 800) / 1024
>>>> which is:
>>>> x * ~500mln
>>>>
>>>> So to be close to overflowing u32 the 'x' has to be > (?=) 8
>>>> (depends on sum_util).
>>>
>>> Sorry but I don't get your point.
>>> This patch is about the return type of compute_energy() and
>>> em_cpu_energy(). And even if we decide to return uW instead of mW,
>>> there is still a lot of margin.
>>>
>>> It's not because you need u64 for computing intermediate value that
>>> you must returns u64
>>
>> The example above shows the need of u64 return value for platforms
>> which are:
>> - 32bit
>> - have e.g. 16 CPUs
>> - has big power value e.g. ~64k mW
>> Then let's to the calc:
>> (64k * 10000) * (16 * 800) / 1024 = ~8000mln = ~8bln
>
> so you return a power consumption of 8kW !!!
>
No. It's in 0.1uW scale, so 800Watts. Which is 16 CPUs * 64Watts
each at max freq and 80% load.
Max power can be < 64Watts, which is 64k milli-Watts (< EM_MAX_POWER)
64k mW * 10000 --> is the 0.1uW precision
On Wed, 7 Jul 2021 at 12:06, Lukasz Luba <lukasz.luba@arm.com> wrote: > > > > On 7/7/21 10:56 AM, Vincent Guittot wrote: > > On Wed, 7 Jul 2021 at 11:48, Lukasz Luba <lukasz.luba@arm.com> wrote: > >> > >> > >> > >> On 7/7/21 10:37 AM, Vincent Guittot wrote: > >>> On Wed, 7 Jul 2021 at 10:23, Lukasz Luba <lukasz.luba@arm.com> wrote: > >>>> > >>>> > >>>> > >>>> On 7/7/21 9:00 AM, Vincent Guittot wrote: > >>>>> On Wed, 7 Jul 2021 at 09:49, Lukasz Luba <lukasz.luba@arm.com> wrote: > >>>>>> > >>>>>> > >>>>>> > >>>>>> On 7/7/21 8:07 AM, Vincent Guittot wrote: > >>>>>>> On Fri, 25 Jun 2021 at 17:26, Lukasz Luba <lukasz.luba@arm.com> wrote: > >>>>>>>> > >>>>>>>> The Energy Aware Scheduler (EAS) tries to find best CPU for a waking up > >>>>>>>> task. It probes many possibilities and compares the estimated energy values > >>>>>>>> for different scenarios. For calculating those energy values it relies on > >>>>>>>> Energy Model (EM) data and em_cpu_energy(). The precision which is used in > >>>>>>>> EM data is in milli-Watts (or abstract scale), which sometimes is not > >>>>>>>> sufficient. In some cases it might happen that two CPUs from different > >>>>>>>> Performance Domains (PDs) get the same calculated value for a given task > >>>>>>>> placement, but in more precised scale, they might differ. This rounding > >>>>>>>> error has to be addressed. This patch prepares EAS code for better > >>>>>>>> precision in the coming EM improvements. > >>>>>>> > >>>>>>> Could you explain why 32bits results are not enough and you need to > >>>>>>> move to 64bits ? > >>>>>>> > >>>>>>> Right now the result is in the range [0..2^32[ mW. If you need more > >>>>>>> precision and you want to return uW instead, you will have a result in > >>>>>>> the range [0..4kW[ which seems to be still enough > >>>>>>> > >>>>>> > >>>>>> Currently we have the max value limit for 'power' in EM which is > >>>>>> EM_MAX_POWER 0xffff (64k - 1). We allow to register such big power > >>>>>> values ~64k mW (~64Watts) for an OPP. Then based on 'power' we > >>>>>> pre-calculate 'cost' fields: > >>>>>> cost[i] = power[i] * freq_max / freq[i] > >>>>>> So, for max freq the cost == power. Let's use that in the example. > >>>>>> > >>>>>> Then the em_cpu_energy() calculates as follow: > >>>>>> cost * sum_util / scale_cpu > >>>>>> We are interested in the first part - the value of multiplication. > >>>>> > >>>>> But all these are internal computations of the energy model. At the > >>>>> end, the computed energy that is returned by compute_energy() and > >>>>> em_cpu_energy(), fits in a long > >>>> > >>>> Let's take a look at existing *10000 precision for x CPUs: > >>>> cost * sum_util / scale_cpu = > >>>> (64k *10000) * (x * 800) / 1024 > >>>> which is: > >>>> x * ~500mln > >>>> > >>>> So to be close to overflowing u32 the 'x' has to be > (?=) 8 > >>>> (depends on sum_util). > >>> > >>> Sorry but I don't get your point. > >>> This patch is about the return type of compute_energy() and > >>> em_cpu_energy(). And even if we decide to return uW instead of mW, > >>> there is still a lot of margin. > >>> > >>> It's not because you need u64 for computing intermediate value that > >>> you must returns u64 > >> > >> The example above shows the need of u64 return value for platforms > >> which are: > >> - 32bit > >> - have e.g. 16 CPUs > >> - has big power value e.g. ~64k mW > >> Then let's to the calc: > >> (64k * 10000) * (16 * 800) / 1024 = ~8000mln = ~8bln > > > > so you return a power consumption of 8kW !!! > > > > No. It's in 0.1uW scale, so 800Watts. Which is 16 CPUs * 64Watts Oh! you want 0.1uW precision .... This doesn't seem realistic at all. I'm not even sure that the power model can even reach an accuracy of 1mW > each at max freq and 80% load. > > Max power can be < 64Watts, which is 64k milli-Watts (< EM_MAX_POWER) > 64k mW * 10000 --> is the 0.1uW precision >
On 7/7/21 11:01 AM, Peter Zijlstra wrote:
> On Wed, Jul 07, 2021 at 09:09:08AM +0100, Lukasz Luba wrote:
>> For now we would live with this simple code which improves
>> all recent 64bit platforms and is easy to take it into Android
>> common kernel. The next step would be more scattered across
>> other subsystems, so harder to backport to Android 5.4 and others.
>
> Ah, you *do* only care about 64bit :-) So one option is to only increase
> precision for 64BIT builds, just like we do for scale_load() and
> friends.
>
Your suggestion is potentially a good compromise :-)
We could leave the 32bit alone and they would have old code and
precision.
Thank you for the comments. Let me discuss this internally with my team.
On 7/7/21 11:11 AM, Vincent Guittot wrote: > On Wed, 7 Jul 2021 at 12:06, Lukasz Luba <lukasz.luba@arm.com> wrote: >> [snip] >> No. It's in 0.1uW scale, so 800Watts. Which is 16 CPUs * 64Watts > > Oh! you want 0.1uW precision .... This doesn't seem realistic at all. > I'm not even sure that the power model can even reach an accuracy of > 1mW > True, the EM is registering platform with 1mW precision, but 1uW precision makes more sense for internal EAS calculation. I don't force platforms to report 1uW power, I just want to operate on it internally. PowerCap and DTPM also operate internally on 1uW, so it's not that unrealistic that some kernel components want better resolution. But as Peter suggested, we might skip 32bit platforms for this issue. I have to discussed this internally.
On Wed, 7 Jul 2021 at 12:29, Lukasz Luba <lukasz.luba@arm.com> wrote: > > > > On 7/7/21 11:11 AM, Vincent Guittot wrote: > > On Wed, 7 Jul 2021 at 12:06, Lukasz Luba <lukasz.luba@arm.com> wrote: > >> > > [snip] > > >> No. It's in 0.1uW scale, so 800Watts. Which is 16 CPUs * 64Watts > > > > Oh! you want 0.1uW precision .... This doesn't seem realistic at all. > > I'm not even sure that the power model can even reach an accuracy of > > 1mW > > > > True, the EM is registering platform with 1mW precision, but 1uW Do you mean 1uW or 0.1uW ? > precision makes more sense for internal EAS calculation. I don't > force platforms to report 1uW power, I just want to operate on > it internally. PowerCap and DTPM also operate internally on 1uW, > so it's not that unrealistic that some kernel components want > better resolution. > > But as Peter suggested, we might skip 32bit platforms for this issue. > I have to discussed this internally.
On 7/7/21 11:32 AM, Vincent Guittot wrote:
> On Wed, 7 Jul 2021 at 12:29, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>
>>
>>
>> On 7/7/21 11:11 AM, Vincent Guittot wrote:
>>> On Wed, 7 Jul 2021 at 12:06, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>
>>
>> [snip]
>>
>>>> No. It's in 0.1uW scale, so 800Watts. Which is 16 CPUs * 64Watts
>>>
>>> Oh! you want 0.1uW precision .... This doesn't seem realistic at all.
>>> I'm not even sure that the power model can even reach an accuracy of
>>> 1mW
>>>
>>
>> True, the EM is registering platform with 1mW precision, but 1uW
>
> Do you mean 1uW or 0.1uW ?
In this patch set I've proposed 0.1uW, but I'm open to drop one
order of magnitude. The 1uW still be good.
On Wed, 7 Jul 2021 at 12:41, Lukasz Luba <lukasz.luba@arm.com> wrote:
>
>
>
> On 7/7/21 11:32 AM, Vincent Guittot wrote:
> > On Wed, 7 Jul 2021 at 12:29, Lukasz Luba <lukasz.luba@arm.com> wrote:
> >>
> >>
> >>
> >> On 7/7/21 11:11 AM, Vincent Guittot wrote:
> >>> On Wed, 7 Jul 2021 at 12:06, Lukasz Luba <lukasz.luba@arm.com> wrote:
> >>>>
> >>
> >> [snip]
> >>
> >>>> No. It's in 0.1uW scale, so 800Watts. Which is 16 CPUs * 64Watts
> >>>
> >>> Oh! you want 0.1uW precision .... This doesn't seem realistic at all.
> >>> I'm not even sure that the power model can even reach an accuracy of
> >>> 1mW
> >>>
> >>
> >> True, the EM is registering platform with 1mW precision, but 1uW
> >
> > Do you mean 1uW or 0.1uW ?
>
> In this patch set I've proposed 0.1uW, but I'm open to drop one
> order of magnitude. The 1uW still be good.
I don't want to underestimate the capabilities of the power model but
I don't see which benefit you will get with 0.1uW precision
With a 1uW precision the long type currently used for the returned
value is fine for 32bits machine AFAICT
On 7/7/21 11:50 AM, Vincent Guittot wrote:
> On Wed, 7 Jul 2021 at 12:41, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>
>>
>>
>> On 7/7/21 11:32 AM, Vincent Guittot wrote:
>>> On Wed, 7 Jul 2021 at 12:29, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>
>>>>
>>>>
>>>> On 7/7/21 11:11 AM, Vincent Guittot wrote:
>>>>> On Wed, 7 Jul 2021 at 12:06, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>>>
>>>>
>>>> [snip]
>>>>
>>>>>> No. It's in 0.1uW scale, so 800Watts. Which is 16 CPUs * 64Watts
>>>>>
>>>>> Oh! you want 0.1uW precision .... This doesn't seem realistic at all.
>>>>> I'm not even sure that the power model can even reach an accuracy of
>>>>> 1mW
>>>>>
>>>>
>>>> True, the EM is registering platform with 1mW precision, but 1uW
>>>
>>> Do you mean 1uW or 0.1uW ?
>>
>> In this patch set I've proposed 0.1uW, but I'm open to drop one
>> order of magnitude. The 1uW still be good.
>
> I don't want to underestimate the capabilities of the power model but
> I don't see which benefit you will get with 0.1uW precision
> With a 1uW precision the long type currently used for the returned
> value is fine for 32bits machine AFAICT
>
For 1uW and 1.2Watts for one core, 4 CPUs in cluster we get:
(1200 * 1000) * (4 * 1024) = ~4.9bln
so it would need div 64 version
On Wed, 7 Jul 2021 at 13:02, Lukasz Luba <lukasz.luba@arm.com> wrote:
>
>
>
> On 7/7/21 11:50 AM, Vincent Guittot wrote:
> > On Wed, 7 Jul 2021 at 12:41, Lukasz Luba <lukasz.luba@arm.com> wrote:
> >>
> >>
> >>
> >> On 7/7/21 11:32 AM, Vincent Guittot wrote:
> >>> On Wed, 7 Jul 2021 at 12:29, Lukasz Luba <lukasz.luba@arm.com> wrote:
> >>>>
> >>>>
> >>>>
> >>>> On 7/7/21 11:11 AM, Vincent Guittot wrote:
> >>>>> On Wed, 7 Jul 2021 at 12:06, Lukasz Luba <lukasz.luba@arm.com> wrote:
> >>>>>>
> >>>>
> >>>> [snip]
> >>>>
> >>>>>> No. It's in 0.1uW scale, so 800Watts. Which is 16 CPUs * 64Watts
> >>>>>
> >>>>> Oh! you want 0.1uW precision .... This doesn't seem realistic at all.
> >>>>> I'm not even sure that the power model can even reach an accuracy of
> >>>>> 1mW
> >>>>>
> >>>>
> >>>> True, the EM is registering platform with 1mW precision, but 1uW
> >>>
> >>> Do you mean 1uW or 0.1uW ?
> >>
> >> In this patch set I've proposed 0.1uW, but I'm open to drop one
> >> order of magnitude. The 1uW still be good.
> >
> > I don't want to underestimate the capabilities of the power model but
> > I don't see which benefit you will get with 0.1uW precision
> > With a 1uW precision the long type currently used for the returned
> > value is fine for 32bits machine AFAICT
> >
>
> For 1uW and 1.2Watts for one core, 4 CPUs in cluster we get:
> (1200 * 1000) * (4 * 1024) = ~4.9bln
> so it would need div 64 version
But as stated before, this is an internal computation step and doesn't
have to be reflected in the returned value which can stay a long
On 7/7/21 2:53 PM, Vincent Guittot wrote:
> On Wed, 7 Jul 2021 at 13:02, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>
>>
>>
>> On 7/7/21 11:50 AM, Vincent Guittot wrote:
>>> On Wed, 7 Jul 2021 at 12:41, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>
>>>>
>>>>
>>>> On 7/7/21 11:32 AM, Vincent Guittot wrote:
>>>>> On Wed, 7 Jul 2021 at 12:29, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 7/7/21 11:11 AM, Vincent Guittot wrote:
>>>>>>> On Wed, 7 Jul 2021 at 12:06, Lukasz Luba <lukasz.luba@arm.com> wrote:
>>>>>>>>
>>>>>>
>>>>>> [snip]
>>>>>>
>>>>>>>> No. It's in 0.1uW scale, so 800Watts. Which is 16 CPUs * 64Watts
>>>>>>>
>>>>>>> Oh! you want 0.1uW precision .... This doesn't seem realistic at all.
>>>>>>> I'm not even sure that the power model can even reach an accuracy of
>>>>>>> 1mW
>>>>>>>
>>>>>>
>>>>>> True, the EM is registering platform with 1mW precision, but 1uW
>>>>>
>>>>> Do you mean 1uW or 0.1uW ?
>>>>
>>>> In this patch set I've proposed 0.1uW, but I'm open to drop one
>>>> order of magnitude. The 1uW still be good.
>>>
>>> I don't want to underestimate the capabilities of the power model but
>>> I don't see which benefit you will get with 0.1uW precision
>>> With a 1uW precision the long type currently used for the returned
>>> value is fine for 32bits machine AFAICT
>>>
>>
>> For 1uW and 1.2Watts for one core, 4 CPUs in cluster we get:
>> (1200 * 1000) * (4 * 1024) = ~4.9bln
>> so it would need div 64 version
>
> But as stated before, this is an internal computation step and doesn't
> have to be reflected in the returned value which can stay a long
>
I agree, we might scale down the result if it's too big, before the
return. We could figure this out at the EM registration point, so a
proper shift might be applied for such platform. It might enable both
32bit and 64bit platforms to avoid the rounding error. Let me experiment
with some code to check all the cases.
Thank you for the comments!