Following the discussion with Qais [1] about how to handle uclamp
requirements and after syncing with him, we agreed that I should move
forward on the patchset to rework the interface between scheduler and
schedutil governor to provide more information to the latter. Scheduler
(and EAS in particular) doesn't need anymore to guess estimate which
headroom the governor wants to apply and will directly ask for the target
freq. Then the governor directly gets the actual utilization and new
minimum and maximum boundaries to select this target frequency and
doesn't have to deal anymore with scheduler internals like uclamp when
including iowait boost.
[1] https://lore.kernel.org/lkml/CAKfTPtA5JqNCauG-rP3wGfq+p8EEVx9Tvwj6ksM3SYCwRmfCTg@mail.gmail.com/
Changes since v1
- fix a bug (always set max even when returning early)
- fix typos
Vincent Guittot (2):
sched/schedutil: rework performance estimation
sched/schedutil: rework iowait boost
include/linux/energy_model.h | 1 -
kernel/sched/core.c | 82 ++++++++++++-------------------
kernel/sched/cpufreq_schedutil.c | 72 +++++++++++++++++----------
kernel/sched/fair.c | 22 +++++++--
kernel/sched/sched.h | 84 +++-----------------------------
5 files changed, 103 insertions(+), 158 deletions(-)
--
2.34.1
The current method to take into account uclamp hints when estimating the
target frequency can end into situation where the selected target
frequency is finally higher than uclamp hints whereas there are no real
needs. Such cases mainly happen because we are currently mixing the
traditional scheduler utilization signal with the uclamp performance
hints. By adding these 2 metrics, we loose an important information when
it comes to select the target frequency and we have to make some
assumptions which can't fit all cases.
Rework the interface between the scheduler and schedutil governor in order
to propagate all information down to the cpufreq governor.
effective_cpu_util() interface changes and now returns the actual
utilization of the CPU with 2 optional inputs:
- The minimum performance for this CPU; typically the capacity to handle
the deadline task and the interrupt pressure. But also uclamp_min
request when available.
- The maximum targeting performance for this CPU which reflects the
maximum level that we would like to not exceed. By default it will be
the CPU capacity but can be reduced because of some performance hints
set with uclamp. The value can be lower than actual utilization and/or
min performance level.
A new sugov_effective_cpu_perf() interface is also available to compute
the final performance level that is targeted for the CPU after applying
some cpufreq headroom and taking into account all inputs.
With these 2 functions, schedutil is now able to decide when it must go
above uclamp hints. It now also have a generic way to get the min
perfromance level.
The dependency between energy model and cpufreq governor and its headroom
policy doesn't exist anymore.
eenv_pd_max_util asks schedutil for the targeted performance after
applying the impact of the waking task.
Signed-off-by: Vincent Guittot <[email protected]>
---
include/linux/energy_model.h | 1 -
kernel/sched/core.c | 82 ++++++++++++--------------------
kernel/sched/cpufreq_schedutil.c | 43 +++++++++++++----
kernel/sched/fair.c | 22 +++++++--
kernel/sched/sched.h | 24 +++-------
5 files changed, 89 insertions(+), 83 deletions(-)
diff --git a/include/linux/energy_model.h b/include/linux/energy_model.h
index b9caa01dfac4..adec808b371a 100644
--- a/include/linux/energy_model.h
+++ b/include/linux/energy_model.h
@@ -243,7 +243,6 @@ static inline unsigned long em_cpu_energy(struct em_perf_domain *pd,
scale_cpu = arch_scale_cpu_capacity(cpu);
ps = &pd->table[pd->nr_perf_states - 1];
- max_util = map_util_perf(max_util);
max_util = min(max_util, allowed_cpu_cap);
freq = map_util_freq(max_util, ps->frequency, scale_cpu);
diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index 7a0c16115b79..af5333327493 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c
@@ -7391,18 +7391,13 @@ int sched_core_idle_cpu(int cpu)
* required to meet deadlines.
*/
unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
- enum cpu_util_type type,
- struct task_struct *p)
+ unsigned long *min,
+ unsigned long *max)
{
- unsigned long dl_util, util, irq, max;
+ unsigned long util, irq, scale;
struct rq *rq = cpu_rq(cpu);
- max = arch_scale_cpu_capacity(cpu);
-
- if (!uclamp_is_used() &&
- type == FREQUENCY_UTIL && rt_rq_is_runnable(&rq->rt)) {
- return max;
- }
+ scale = arch_scale_cpu_capacity(cpu);
/*
* Early check to see if IRQ/steal time saturates the CPU, can be
@@ -7410,45 +7405,41 @@ unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
* update_irq_load_avg().
*/
irq = cpu_util_irq(rq);
- if (unlikely(irq >= max))
- return max;
+ if (unlikely(irq >= scale)) {
+ if (min)
+ *min = scale;
+ if (max)
+ *max = scale;
+ return scale;
+ }
+
+ /*
+ * The minimum utilization returns the highest level between:
+ * - the computed DL bandwidth needed with the irq pressure which
+ * steals time to the deadline task.
+ * - The minimum bandwidth requirement for CFS and/or RT.
+ */
+ if (min)
+ *min = max(irq + cpu_bw_dl(rq), uclamp_rq_get(rq, UCLAMP_MIN));
/*
* Because the time spend on RT/DL tasks is visible as 'lost' time to
* CFS tasks and we use the same metric to track the effective
* utilization (PELT windows are synchronized) we can directly add them
* to obtain the CPU's actual utilization.
- *
- * CFS and RT utilization can be boosted or capped, depending on
- * utilization clamp constraints requested by currently RUNNABLE
- * tasks.
- * When there are no CFS RUNNABLE tasks, clamps are released and
- * frequency will be gracefully reduced with the utilization decay.
*/
util = util_cfs + cpu_util_rt(rq);
- if (type == FREQUENCY_UTIL)
- util = uclamp_rq_util_with(rq, util, p);
-
- dl_util = cpu_util_dl(rq);
+ util += cpu_util_dl(rq);
/*
- * For frequency selection we do not make cpu_util_dl() a permanent part
- * of this sum because we want to use cpu_bw_dl() later on, but we need
- * to check if the CFS+RT+DL sum is saturated (ie. no idle time) such
- * that we select f_max when there is no idle time.
- *
- * NOTE: numerical errors or stop class might cause us to not quite hit
- * saturation when we should -- something for later.
+ * The maximum hint is a soft bandwidth requirement which can be lower
+ * than the actual utilization because of uclamp_max requirements
*/
- if (util + dl_util >= max)
- return max;
+ if (max)
+ *max = min(scale, uclamp_rq_get(rq, UCLAMP_MAX));
- /*
- * OTOH, for energy computation we need the estimated running time, so
- * include util_dl and ignore dl_bw.
- */
- if (type == ENERGY_UTIL)
- util += dl_util;
+ if (util >= scale)
+ return scale;
/*
* There is still idle time; further improve the number by using the
@@ -7459,28 +7450,15 @@ unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
* U' = irq + --------- * U
* max
*/
- util = scale_irq_capacity(util, irq, max);
+ util = scale_irq_capacity(util, irq, scale);
util += irq;
- /*
- * Bandwidth required by DEADLINE must always be granted while, for
- * FAIR and RT, we use blocked utilization of IDLE CPUs as a mechanism
- * to gracefully reduce the frequency when no tasks show up for longer
- * periods of time.
- *
- * Ideally we would like to set bw_dl as min/guaranteed freq and util +
- * bw_dl as requested freq. However, cpufreq is not yet ready for such
- * an interface. So, we only do the latter for now.
- */
- if (type == FREQUENCY_UTIL)
- util += cpu_bw_dl(rq);
-
- return min(max, util);
+ return min(scale, util);
}
unsigned long sched_cpu_util(int cpu)
{
- return effective_cpu_util(cpu, cpu_util_cfs(cpu), ENERGY_UTIL, NULL);
+ return effective_cpu_util(cpu, cpu_util_cfs(cpu), NULL, NULL);
}
#endif /* CONFIG_SMP */
diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
index 458d359f5991..be90d7ac0140 100644
--- a/kernel/sched/cpufreq_schedutil.c
+++ b/kernel/sched/cpufreq_schedutil.c
@@ -47,7 +47,7 @@ struct sugov_cpu {
u64 last_update;
unsigned long util;
- unsigned long bw_dl;
+ unsigned long bw_min;
/* The field below is for single-CPU policies only: */
#ifdef CONFIG_NO_HZ_COMMON
@@ -143,7 +143,6 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy,
unsigned int freq = arch_scale_freq_invariant() ?
policy->cpuinfo.max_freq : policy->cur;
- util = map_util_perf(util);
freq = map_util_freq(util, freq, max);
if (freq == sg_policy->cached_raw_freq && !sg_policy->need_freq_update)
@@ -153,14 +152,38 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy,
return cpufreq_driver_resolve_freq(policy, freq);
}
+unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
+ unsigned long min,
+ unsigned long max)
+{
+ unsigned long target;
+ struct rq *rq = cpu_rq(cpu);
+
+ if (rt_rq_is_runnable(&rq->rt))
+ return max;
+
+ /* Provide at least enough capacity for DL + IRQ */
+ target = min;
+
+ actual = map_util_perf(actual);
+ /* Actually we don't need to target the max performance */
+ if (actual < max)
+ max = actual;
+
+ /*
+ * Ensure at least minimum performance while providing more compute
+ * capacity when possible.
+ */
+ return max(target, max);
+}
+
static void sugov_get_util(struct sugov_cpu *sg_cpu)
{
- unsigned long util = cpu_util_cfs_boost(sg_cpu->cpu);
- struct rq *rq = cpu_rq(sg_cpu->cpu);
+ unsigned long min, max, util = cpu_util_cfs_boost(sg_cpu->cpu);
- sg_cpu->bw_dl = cpu_bw_dl(rq);
- sg_cpu->util = effective_cpu_util(sg_cpu->cpu, util,
- FREQUENCY_UTIL, NULL);
+ util = effective_cpu_util(sg_cpu->cpu, util, &min, &max);
+ sg_cpu->bw_min = map_util_perf(min);
+ sg_cpu->util = sugov_effective_cpu_perf(sg_cpu->cpu, util, min, max);
}
/**
@@ -306,7 +329,7 @@ static inline bool sugov_cpu_is_busy(struct sugov_cpu *sg_cpu) { return false; }
*/
static inline void ignore_dl_rate_limit(struct sugov_cpu *sg_cpu)
{
- if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_dl)
+ if (cpu_bw_dl(cpu_rq(sg_cpu->cpu)) > sg_cpu->bw_min)
sg_cpu->sg_policy->limits_changed = true;
}
@@ -407,8 +430,8 @@ static void sugov_update_single_perf(struct update_util_data *hook, u64 time,
sugov_cpu_is_busy(sg_cpu) && sg_cpu->util < prev_util)
sg_cpu->util = prev_util;
- cpufreq_driver_adjust_perf(sg_cpu->cpu, map_util_perf(sg_cpu->bw_dl),
- map_util_perf(sg_cpu->util), max_cap);
+ cpufreq_driver_adjust_perf(sg_cpu->cpu, sg_cpu->bw_min,
+ sg_cpu->util, max_cap);
sg_cpu->sg_policy->last_freq_update_time = time;
}
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 8767988242ee..ed64f2eaaa2a 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -7671,7 +7671,7 @@ static inline void eenv_pd_busy_time(struct energy_env *eenv,
for_each_cpu(cpu, pd_cpus) {
unsigned long util = cpu_util(cpu, p, -1, 0);
- busy_time += effective_cpu_util(cpu, util, ENERGY_UTIL, NULL);
+ busy_time += effective_cpu_util(cpu, util, NULL, NULL);
}
eenv->pd_busy_time = min(eenv->pd_cap, busy_time);
@@ -7694,7 +7694,7 @@ eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
for_each_cpu(cpu, pd_cpus) {
struct task_struct *tsk = (cpu == dst_cpu) ? p : NULL;
unsigned long util = cpu_util(cpu, p, dst_cpu, 1);
- unsigned long eff_util;
+ unsigned long eff_util, min, max;
/*
* Performance domain frequency: utilization clamping
@@ -7703,7 +7703,23 @@ eenv_pd_max_util(struct energy_env *eenv, struct cpumask *pd_cpus,
* NOTE: in case RT tasks are running, by default the
* FREQUENCY_UTIL's utilization can be max OPP.
*/
- eff_util = effective_cpu_util(cpu, util, FREQUENCY_UTIL, tsk);
+ eff_util = effective_cpu_util(cpu, util, &min, &max);
+
+ /* Task's uclamp can modify min and max value */
+ if (tsk && uclamp_is_used()) {
+ min = max(min, uclamp_eff_value(p, UCLAMP_MIN));
+
+ /*
+ * If there is no active max uclamp constraint,
+ * directly use task's one otherwise keep max
+ */
+ if (uclamp_rq_is_idle(cpu_rq(cpu)))
+ max = uclamp_eff_value(p, UCLAMP_MAX);
+ else
+ max = max(max, uclamp_eff_value(p, UCLAMP_MAX));
+ }
+
+ eff_util = sugov_effective_cpu_perf(cpu, eff_util, min, max);
max_util = max(max_util, eff_util);
}
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 2e5a95486a42..302b451a3fd8 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -2961,24 +2961,14 @@ static inline void cpufreq_update_util(struct rq *rq, unsigned int flags) {}
#endif
#ifdef CONFIG_SMP
-/**
- * enum cpu_util_type - CPU utilization type
- * @FREQUENCY_UTIL: Utilization used to select frequency
- * @ENERGY_UTIL: Utilization used during energy calculation
- *
- * The utilization signals of all scheduling classes (CFS/RT/DL) and IRQ time
- * need to be aggregated differently depending on the usage made of them. This
- * enum is used within effective_cpu_util() to differentiate the types of
- * utilization expected by the callers, and adjust the aggregation accordingly.
- */
-enum cpu_util_type {
- FREQUENCY_UTIL,
- ENERGY_UTIL,
-};
-
unsigned long effective_cpu_util(int cpu, unsigned long util_cfs,
- enum cpu_util_type type,
- struct task_struct *p);
+ unsigned long *min,
+ unsigned long *max);
+
+unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
+ unsigned long min,
+ unsigned long max);
+
/*
* Verify the fitness of task @p to run on @cpu taking into account the
--
2.34.1
Use the max value that has already been computed inside sugov_get_util()
to cap the iowait boost and remove dependency with uclamp_rq_util_with()
which is not used anymore.
Signed-off-by: Vincent Guittot <[email protected]>
---
kernel/sched/cpufreq_schedutil.c | 29 ++++++++-------
kernel/sched/sched.h | 60 --------------------------------
2 files changed, 14 insertions(+), 75 deletions(-)
diff --git a/kernel/sched/cpufreq_schedutil.c b/kernel/sched/cpufreq_schedutil.c
index be90d7ac0140..efd4b77471ab 100644
--- a/kernel/sched/cpufreq_schedutil.c
+++ b/kernel/sched/cpufreq_schedutil.c
@@ -177,11 +177,12 @@ unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
return max(target, max);
}
-static void sugov_get_util(struct sugov_cpu *sg_cpu)
+static void sugov_get_util(struct sugov_cpu *sg_cpu, unsigned long boost)
{
unsigned long min, max, util = cpu_util_cfs_boost(sg_cpu->cpu);
util = effective_cpu_util(sg_cpu->cpu, util, &min, &max);
+ util = max(util, boost);
sg_cpu->bw_min = map_util_perf(min);
sg_cpu->util = sugov_effective_cpu_perf(sg_cpu->cpu, util, min, max);
}
@@ -274,18 +275,16 @@ static void sugov_iowait_boost(struct sugov_cpu *sg_cpu, u64 time,
* This mechanism is designed to boost high frequently IO waiting tasks, while
* being more conservative on tasks which does sporadic IO operations.
*/
-static void sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
+static unsigned long sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
unsigned long max_cap)
{
- unsigned long boost;
-
/* No boost currently required */
if (!sg_cpu->iowait_boost)
- return;
+ return 0;
/* Reset boost if the CPU appears to have been idle enough */
if (sugov_iowait_reset(sg_cpu, time, false))
- return;
+ return 0;
if (!sg_cpu->iowait_boost_pending) {
/*
@@ -294,7 +293,7 @@ static void sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
sg_cpu->iowait_boost >>= 1;
if (sg_cpu->iowait_boost < IOWAIT_BOOST_MIN) {
sg_cpu->iowait_boost = 0;
- return;
+ return 0;
}
}
@@ -304,10 +303,7 @@ static void sugov_iowait_apply(struct sugov_cpu *sg_cpu, u64 time,
* sg_cpu->util is already in capacity scale; convert iowait_boost
* into the same scale so we can compare.
*/
- boost = (sg_cpu->iowait_boost * max_cap) >> SCHED_CAPACITY_SHIFT;
- boost = uclamp_rq_util_with(cpu_rq(sg_cpu->cpu), boost, NULL);
- if (sg_cpu->util < boost)
- sg_cpu->util = boost;
+ return (sg_cpu->iowait_boost * max_cap) >> SCHED_CAPACITY_SHIFT;
}
#ifdef CONFIG_NO_HZ_COMMON
@@ -337,6 +333,8 @@ static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
u64 time, unsigned long max_cap,
unsigned int flags)
{
+ unsigned long boost;
+
sugov_iowait_boost(sg_cpu, time, flags);
sg_cpu->last_update = time;
@@ -345,8 +343,8 @@ static inline bool sugov_update_single_common(struct sugov_cpu *sg_cpu,
if (!sugov_should_update_freq(sg_cpu->sg_policy, time))
return false;
- sugov_get_util(sg_cpu);
- sugov_iowait_apply(sg_cpu, time, max_cap);
+ boost = sugov_iowait_apply(sg_cpu, time, max_cap);
+ sugov_get_util(sg_cpu, boost);
return true;
}
@@ -447,9 +445,10 @@ static unsigned int sugov_next_freq_shared(struct sugov_cpu *sg_cpu, u64 time)
for_each_cpu(j, policy->cpus) {
struct sugov_cpu *j_sg_cpu = &per_cpu(sugov_cpu, j);
+ unsigned long boost;
- sugov_get_util(j_sg_cpu);
- sugov_iowait_apply(j_sg_cpu, time, max_cap);
+ boost = sugov_iowait_apply(j_sg_cpu, time, max_cap);
+ sugov_get_util(j_sg_cpu, boost);
util = max(j_sg_cpu->util, util);
}
diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index 302b451a3fd8..e3cb8e004bd1 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h
@@ -3025,59 +3025,6 @@ static inline bool uclamp_rq_is_idle(struct rq *rq)
return rq->uclamp_flags & UCLAMP_FLAG_IDLE;
}
-/**
- * uclamp_rq_util_with - clamp @util with @rq and @p effective uclamp values.
- * @rq: The rq to clamp against. Must not be NULL.
- * @util: The util value to clamp.
- * @p: The task to clamp against. Can be NULL if you want to clamp
- * against @rq only.
- *
- * Clamps the passed @util to the max(@rq, @p) effective uclamp values.
- *
- * If sched_uclamp_used static key is disabled, then just return the util
- * without any clamping since uclamp aggregation at the rq level in the fast
- * path is disabled, rendering this operation a NOP.
- *
- * Use uclamp_eff_value() if you don't care about uclamp values at rq level. It
- * will return the correct effective uclamp value of the task even if the
- * static key is disabled.
- */
-static __always_inline
-unsigned long uclamp_rq_util_with(struct rq *rq, unsigned long util,
- struct task_struct *p)
-{
- unsigned long min_util = 0;
- unsigned long max_util = 0;
-
- if (!static_branch_likely(&sched_uclamp_used))
- return util;
-
- if (p) {
- min_util = uclamp_eff_value(p, UCLAMP_MIN);
- max_util = uclamp_eff_value(p, UCLAMP_MAX);
-
- /*
- * Ignore last runnable task's max clamp, as this task will
- * reset it. Similarly, no need to read the rq's min clamp.
- */
- if (uclamp_rq_is_idle(rq))
- goto out;
- }
-
- min_util = max_t(unsigned long, min_util, uclamp_rq_get(rq, UCLAMP_MIN));
- max_util = max_t(unsigned long, max_util, uclamp_rq_get(rq, UCLAMP_MAX));
-out:
- /*
- * Since CPU's {min,max}_util clamps are MAX aggregated considering
- * RUNNABLE tasks with _different_ clamps, we can end up with an
- * inversion. Fix it now when the clamps are applied.
- */
- if (unlikely(min_util >= max_util))
- return min_util;
-
- return clamp(util, min_util, max_util);
-}
-
/* Is the rq being capped/throttled by uclamp_max? */
static inline bool uclamp_rq_is_capped(struct rq *rq)
{
@@ -3115,13 +3062,6 @@ static inline unsigned long uclamp_eff_value(struct task_struct *p,
return SCHED_CAPACITY_SCALE;
}
-static inline
-unsigned long uclamp_rq_util_with(struct rq *rq, unsigned long util,
- struct task_struct *p)
-{
- return util;
-}
-
static inline bool uclamp_rq_is_capped(struct rq *rq) { return false; }
static inline bool uclamp_is_used(void)
--
2.34.1
Hi Vincent,
On 10/26/23 18:09, Vincent Guittot wrote:
> The current method to take into account uclamp hints when estimating the
> target frequency can end into situation where the selected target
> frequency is finally higher than uclamp hints whereas there are no real
> needs. Such cases mainly happen because we are currently mixing the
> traditional scheduler utilization signal with the uclamp performance
> hints. By adding these 2 metrics, we loose an important information when
> it comes to select the target frequency and we have to make some
> assumptions which can't fit all cases.
>
> Rework the interface between the scheduler and schedutil governor in order
> to propagate all information down to the cpufreq governor.
>
> effective_cpu_util() interface changes and now returns the actual
> utilization of the CPU with 2 optional inputs:
> - The minimum performance for this CPU; typically the capacity to handle
> the deadline task and the interrupt pressure. But also uclamp_min
> request when available.
> - The maximum targeting performance for this CPU which reflects the
> maximum level that we would like to not exceed. By default it will be
> the CPU capacity but can be reduced because of some performance hints
> set with uclamp. The value can be lower than actual utilization and/or
> min performance level.
You have probably missed my question in the last v1 patch set.
The description above needs a bit of clarification, since looking at the
patches some dark corners are introduced IMO:
Currently, we have a less aggressive power saving policy than this
proposal.
The questions:
What if the PD has 4 CPUs, the max util found is 500 and is from a CPU
w/ uclamp_max, but there is another CPU with normal utilization 499?
What should be the final frequency for that PD?
In current design, where we care more about 'delivered performance
to the tasks' than power saving, the +20% would be applied for the
frequency. Therefore if that CPU with 499 util doesn't have uclamp_max,
it would get a decent amount of idle time for its tasks (to compensate
some workload variation).
Regards,
Lukasz
Hi Lukasz,
On Mon, 30 Oct 2023 at 18:45, Lukasz Luba <[email protected]> wrote:
>
> Hi Vincent,
>
> On 10/26/23 18:09, Vincent Guittot wrote:
> > The current method to take into account uclamp hints when estimating the
> > target frequency can end into situation where the selected target
> > frequency is finally higher than uclamp hints whereas there are no real
> > needs. Such cases mainly happen because we are currently mixing the
> > traditional scheduler utilization signal with the uclamp performance
> > hints. By adding these 2 metrics, we loose an important information when
> > it comes to select the target frequency and we have to make some
> > assumptions which can't fit all cases.
> >
> > Rework the interface between the scheduler and schedutil governor in order
> > to propagate all information down to the cpufreq governor.
> >
> > effective_cpu_util() interface changes and now returns the actual
> > utilization of the CPU with 2 optional inputs:
> > - The minimum performance for this CPU; typically the capacity to handle
> > the deadline task and the interrupt pressure. But also uclamp_min
> > request when available.
> > - The maximum targeting performance for this CPU which reflects the
> > maximum level that we would like to not exceed. By default it will be
> > the CPU capacity but can be reduced because of some performance hints
> > set with uclamp. The value can be lower than actual utilization and/or
> > min performance level.
>
> You have probably missed my question in the last v1 patch set.
Yes, sorry
>
> The description above needs a bit of clarification, since looking at the
> patches some dark corners are introduced IMO:
>
> Currently, we have a less aggressive power saving policy than this
> proposal.
>
> The questions:
> What if the PD has 4 CPUs, the max util found is 500 and is from a CPU
> w/ uclamp_max, but there is another CPU with normal utilization 499?
> What should be the final frequency for that PD?
We now follow the same sequence everywhere which can be summarized by:
for each cpu sharing the same frequency domain:
util = cpu_util(cpu)
eff_util = effective_cpu_util(util, &min, &max)
eff_util = sugov_effective_cpu_perf(eff_util, min, max) which
applies the dvfs headroom if needed
max_util = max(max_util, eff_util);
EAS anticipates the impact of the waking task on utilization and max
but the end result is the same as above once the task is enqueued so I
didn't show it for simplicity
Coming back to your example
CPU0 has uclamp_max = 500 and an actual utilization above 500. Its
eff_util will be 500
CPU1 doesn't have uclamp_max constraint and an actual utilization of
499 which will be increase with dvfs headroom to 623 in
sugov_effective_cpu_perf()
The final max util will be 623
With the current implementation we apply the dvfs headroom to the
final max_util (which is the CPU0 with uclamp_max == 500) whereas we
now apply the dvfs headroom on each CPU inside
sugov_effective_cpu_perf()
The main difference is that if CPU1 has an actual utilization of 400,
the max_util of the frequency domain will be 500 whereas it is 625
after applying dvfs headroom with current implementation
>
> In current design, where we care more about 'delivered performance
> to the tasks' than power saving, the +20% would be applied for the
> frequency. Therefore if that CPU with 499 util doesn't have uclamp_max,
> it would get a decent amount of idle time for its tasks (to compensate
> some workload variation).
CPU1 with 499 still gets its 25% margin or I missed something in your example ?
Vincent
>
> Regards,
> Lukasz
On 26/10/2023 19:09, Vincent Guittot wrote:
[...]
> @@ -153,14 +152,38 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy,
> return cpufreq_driver_resolve_freq(policy, freq);
> }
>
> +unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
> + unsigned long min,
> + unsigned long max)
> +{
> + unsigned long target;
> + struct rq *rq = cpu_rq(cpu);
> +
> + if (rt_rq_is_runnable(&rq->rt))
> + return max;
> +
> + /* Provide at least enough capacity for DL + IRQ */
> + target = min;
> +
> + actual = map_util_perf(actual);
> + /* Actually we don't need to target the max performance */
> + if (actual < max)
> + max = actual;
> +
> + /*
> + * Ensure at least minimum performance while providing more compute
> + * capacity when possible.
> + */
> + return max(target, max);
The superfluous `unsigned long target` is still there?
return max(min, max) is much cleaer.
[...]
On Thu, 2 Nov 2023 at 13:03, Dietmar Eggemann <[email protected]> wrote:
>
> On 26/10/2023 19:09, Vincent Guittot wrote:
>
> [...]
>
> > @@ -153,14 +152,38 @@ static unsigned int get_next_freq(struct sugov_policy *sg_policy,
> > return cpufreq_driver_resolve_freq(policy, freq);
> > }
> >
> > +unsigned long sugov_effective_cpu_perf(int cpu, unsigned long actual,
> > + unsigned long min,
> > + unsigned long max)
> > +{
> > + unsigned long target;
> > + struct rq *rq = cpu_rq(cpu);
> > +
> > + if (rt_rq_is_runnable(&rq->rt))
> > + return max;
> > +
> > + /* Provide at least enough capacity for DL + IRQ */
> > + target = min;
> > +
> > + actual = map_util_perf(actual);
> > + /* Actually we don't need to target the max performance */
> > + if (actual < max)
> > + max = actual;
> > +
> > + /*
> > + * Ensure at least minimum performance while providing more compute
> > + * capacity when possible.
> > + */
> > + return max(target, max);
>
> The superfluous `unsigned long target` is still there?
Yeah, I forgot to add the change in this version but it's in the next one
>
> return max(min, max) is much cleaer.
>
> [...]
Hi Vincent,
I know that there is v3, but just to respond to this below.
On 10/31/23 09:48, Vincent Guittot wrote:
> Hi Lukasz,
>
> On Mon, 30 Oct 2023 at 18:45, Lukasz Luba <[email protected]> wrote:
>>
>> Hi Vincent,
>>
>> On 10/26/23 18:09, Vincent Guittot wrote:
>>> The current method to take into account uclamp hints when estimating the
>>> target frequency can end into situation where the selected target
>>> frequency is finally higher than uclamp hints whereas there are no real
>>> needs. Such cases mainly happen because we are currently mixing the
>>> traditional scheduler utilization signal with the uclamp performance
>>> hints. By adding these 2 metrics, we loose an important information when
>>> it comes to select the target frequency and we have to make some
>>> assumptions which can't fit all cases.
>>>
>>> Rework the interface between the scheduler and schedutil governor in order
>>> to propagate all information down to the cpufreq governor.
>>>
>>> effective_cpu_util() interface changes and now returns the actual
>>> utilization of the CPU with 2 optional inputs:
>>> - The minimum performance for this CPU; typically the capacity to handle
>>> the deadline task and the interrupt pressure. But also uclamp_min
>>> request when available.
>>> - The maximum targeting performance for this CPU which reflects the
>>> maximum level that we would like to not exceed. By default it will be
>>> the CPU capacity but can be reduced because of some performance hints
>>> set with uclamp. The value can be lower than actual utilization and/or
>>> min performance level.
>>
>> You have probably missed my question in the last v1 patch set.
>
> Yes, sorry
>
>>
>> The description above needs a bit of clarification, since looking at the
>> patches some dark corners are introduced IMO:
>>
>> Currently, we have a less aggressive power saving policy than this
>> proposal.
>>
>> The questions:
>> What if the PD has 4 CPUs, the max util found is 500 and is from a CPU
>> w/ uclamp_max, but there is another CPU with normal utilization 499?
>> What should be the final frequency for that PD?
>
> We now follow the same sequence everywhere which can be summarized by:
>
> for each cpu sharing the same frequency domain:
> util = cpu_util(cpu)
> eff_util = effective_cpu_util(util, &min, &max)
> eff_util = sugov_effective_cpu_perf(eff_util, min, max) which
> applies the dvfs headroom if needed
> max_util = max(max_util, eff_util);
>
> EAS anticipates the impact of the waking task on utilization and max
> but the end result is the same as above once the task is enqueued so I
> didn't show it for simplicity
>
> Coming back to your example
> CPU0 has uclamp_max = 500 and an actual utilization above 500. Its
> eff_util will be 500
> CPU1 doesn't have uclamp_max constraint and an actual utilization of
> 499 which will be increase with dvfs headroom to 623 in
> sugov_effective_cpu_perf()
>
> The final max util will be 623
>
> With the current implementation we apply the dvfs headroom to the
> final max_util (which is the CPU0 with uclamp_max == 500) whereas we
> now apply the dvfs headroom on each CPU inside
> sugov_effective_cpu_perf()
>
> The main difference is that if CPU1 has an actual utilization of 400,
> the max_util of the frequency domain will be 500 whereas it is 625
> after applying dvfs headroom with current implementation
>
>>
>> In current design, where we care more about 'delivered performance
>> to the tasks' than power saving, the +20% would be applied for the
>> frequency. Therefore if that CPU with 499 util doesn't have uclamp_max,
>> it would get a decent amount of idle time for its tasks (to compensate
>> some workload variation).
>
> CPU1 with 499 still gets its 25% margin or I missed something in your example ?
You understood this correctly. I don't have more questions. It should
than work OK.
Thanks,
Lukasz