linux/kernel/sched/pelt.c

397 lines
10 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0
/*
* Per Entity Load Tracking
*
* Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
*
* Interactivity improvements by Mike Galbraith
* (C) 2007 Mike Galbraith <efault@gmx.de>
*
* Various enhancements by Dmitry Adamushko.
* (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
*
* Group scheduling enhancements by Srivatsa Vaddagiri
* Copyright IBM Corporation, 2007
* Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
*
* Scaled math optimizations by Thomas Gleixner
* Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
*
* Adaptive scheduling granularity, math enhancements by Peter Zijlstra
* Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra
*
* Move PELT related code from fair.c into this pelt.c file
* Author: Vincent Guittot <vincent.guittot@linaro.org>
*/
#include <linux/sched.h>
#include "sched.h"
#include "pelt.h"
/*
* Approximate:
* val * y^n, where y^32 ~= 0.5 (~1 scheduling period)
*/
static u64 decay_load(u64 val, u64 n)
{
unsigned int local_n;
if (unlikely(n > LOAD_AVG_PERIOD * 63))
return 0;
/* after bounds checking we can collapse to 32-bit */
local_n = n;
/*
* As y^PERIOD = 1/2, we can combine
* y^n = 1/2^(n/PERIOD) * y^(n%PERIOD)
* With a look-up table which covers y^n (n<PERIOD)
*
* To achieve constant time decay_load.
*/
if (unlikely(local_n >= LOAD_AVG_PERIOD)) {
val >>= local_n / LOAD_AVG_PERIOD;
local_n %= LOAD_AVG_PERIOD;
}
val = mul_u64_u32_shr(val, runnable_avg_yN_inv[local_n], 32);
return val;
}
static u32 __accumulate_pelt_segments(u64 periods, u32 d1, u32 d3)
{
u32 c1, c2, c3 = d3; /* y^0 == 1 */
/*
* c1 = d1 y^p
*/
c1 = decay_load((u64)d1, periods);
/*
* p-1
* c2 = 1024 \Sum y^n
* n=1
*
* inf inf
* = 1024 ( \Sum y^n - \Sum y^n - y^0 )
* n=0 n=p
*/
c2 = LOAD_AVG_MAX - decay_load(LOAD_AVG_MAX, periods) - 1024;
return c1 + c2 + c3;
}
#define cap_scale(v, s) ((v)*(s) >> SCHED_CAPACITY_SHIFT)
/*
* Accumulate the three separate parts of the sum; d1 the remainder
* of the last (incomplete) period, d2 the span of full periods and d3
* the remainder of the (incomplete) current period.
*
* d1 d2 d3
* ^ ^ ^
* | | |
* |<->|<----------------->|<--->|
* ... |---x---|------| ... |------|-----x (now)
*
* p-1
* u' = (u + d1) y^p + 1024 \Sum y^n + d3 y^0
* n=1
*
* = u y^p + (Step 1)
*
* p-1
* d1 y^p + 1024 \Sum y^n + d3 y^0 (Step 2)
* n=1
*/
static __always_inline u32
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
accumulate_sum(u64 delta, struct sched_avg *sa,
unsigned long load, unsigned long runnable, int running)
{
u32 contrib = (u32)delta; /* p == 0 -> delta < 1024 */
u64 periods;
delta += sa->period_contrib;
periods = delta / 1024; /* A period is 1024us (~1ms) */
/*
* Step 1: decay old *_sum if we crossed period boundaries.
*/
if (periods) {
sa->load_sum = decay_load(sa->load_sum, periods);
sa->runnable_load_sum =
decay_load(sa->runnable_load_sum, periods);
sa->util_sum = decay_load((u64)(sa->util_sum), periods);
/*
* Step 2
*/
delta %= 1024;
contrib = __accumulate_pelt_segments(periods,
1024 - sa->period_contrib, delta);
}
sa->period_contrib = delta;
if (load)
sa->load_sum += load * contrib;
if (runnable)
sa->runnable_load_sum += runnable * contrib;
if (running)
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
sa->util_sum += contrib << SCHED_CAPACITY_SHIFT;
return periods;
}
/*
* We can represent the historical contribution to runnable average as the
* coefficients of a geometric series. To do this we sub-divide our runnable
* history into segments of approximately 1ms (1024us); label the segment that
* occurred N-ms ago p_N, with p_0 corresponding to the current period, e.g.
*
* [<- 1024us ->|<- 1024us ->|<- 1024us ->| ...
* p0 p1 p2
* (now) (~1ms ago) (~2ms ago)
*
* Let u_i denote the fraction of p_i that the entity was runnable.
*
* We then designate the fractions u_i as our co-efficients, yielding the
* following representation of historical load:
* u_0 + u_1*y + u_2*y^2 + u_3*y^3 + ...
*
* We choose y based on the with of a reasonably scheduling period, fixing:
* y^32 = 0.5
*
* This means that the contribution to load ~32ms ago (u_32) will be weighted
* approximately half as much as the contribution to load within the last ms
* (u_0).
*
* When a period "rolls over" and we have new u_0`, multiplying the previous
* sum again by y is sufficient to update:
* load_avg = u_0` + y*(u_0 + u_1*y + u_2*y^2 + ... )
* = u_0 + u_1*y + u_2*y^2 + ... [re-labeling u_i --> u_{i+1}]
*/
static __always_inline int
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
___update_load_sum(u64 now, struct sched_avg *sa,
unsigned long load, unsigned long runnable, int running)
{
u64 delta;
delta = now - sa->last_update_time;
/*
* This should only happen when time goes backwards, which it
* unfortunately does during sched clock init when we swap over to TSC.
*/
if ((s64)delta < 0) {
sa->last_update_time = now;
return 0;
}
/*
* Use 1024ns as the unit of measurement since it's a reasonable
* approximation of 1us and fast to compute.
*/
delta >>= 10;
if (!delta)
return 0;
sa->last_update_time += delta << 10;
/*
* running is a subset of runnable (weight) so running can't be set if
* runnable is clear. But there are some corner cases where the current
* se has been already dequeued but cfs_rq->curr still points to it.
* This means that weight will be 0 but not running for a sched_entity
* but also for a cfs_rq if the latter becomes idle. As an example,
* this happens during idle_balance() which calls
* update_blocked_averages()
*/
if (!load)
runnable = running = 0;
/*
* Now we know we crossed measurement unit boundaries. The *_avg
* accrues by two steps:
*
* Step 1: accumulate *_sum since last_update_time. If we haven't
* crossed period boundaries, finish.
*/
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
if (!accumulate_sum(delta, sa, load, runnable, running))
return 0;
return 1;
}
static __always_inline void
___update_load_avg(struct sched_avg *sa, unsigned long load, unsigned long runnable)
{
u32 divider = LOAD_AVG_MAX - 1024 + sa->period_contrib;
/*
* Step 2: update *_avg.
*/
sa->load_avg = div_u64(load * sa->load_sum, divider);
sa->runnable_load_avg = div_u64(runnable * sa->runnable_load_sum, divider);
WRITE_ONCE(sa->util_avg, sa->util_sum / divider);
}
/*
* sched_entity:
*
* task:
* se_runnable() == se_weight()
*
* group: [ see update_cfs_group() ]
* se_weight() = tg->weight * grq->load_avg / tg->load_avg
* se_runnable() = se_weight(se) * grq->runnable_load_avg / grq->load_avg
*
* load_sum := runnable_sum
* load_avg = se_weight(se) * runnable_avg
*
* runnable_load_sum := runnable_sum
* runnable_load_avg = se_runnable(se) * runnable_avg
*
* XXX collapse load_sum and runnable_load_sum
*
* cfq_rq:
*
* load_sum = \Sum se_weight(se) * se->avg.load_sum
* load_avg = \Sum se->avg.load_avg
*
* runnable_load_sum = \Sum se_runnable(se) * se->avg.runnable_load_sum
* runnable_load_avg = \Sum se->avg.runable_load_avg
*/
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
int __update_load_avg_blocked_se(u64 now, struct sched_entity *se)
{
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
if (___update_load_sum(now, &se->avg, 0, 0, 0)) {
___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
return 1;
}
return 0;
}
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
int __update_load_avg_se(u64 now, struct cfs_rq *cfs_rq, struct sched_entity *se)
{
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
if (___update_load_sum(now, &se->avg, !!se->on_rq, !!se->on_rq,
cfs_rq->curr == se)) {
___update_load_avg(&se->avg, se_weight(se), se_runnable(se));
cfs_se_util_change(&se->avg);
return 1;
}
return 0;
}
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
int __update_load_avg_cfs_rq(u64 now, struct cfs_rq *cfs_rq)
{
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
if (___update_load_sum(now, &cfs_rq->avg,
scale_load_down(cfs_rq->load.weight),
scale_load_down(cfs_rq->runnable_weight),
cfs_rq->curr != NULL)) {
___update_load_avg(&cfs_rq->avg, 1, 1);
return 1;
}
return 0;
}
2018-06-28 15:45:05 +00:00
/*
* rt_rq:
*
* util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
* util_sum = cpu_scale * load_sum
* runnable_load_sum = load_sum
*
* load_avg and runnable_load_avg are not supported and meaningless.
*
*/
int update_rt_rq_load_avg(u64 now, struct rq *rq, int running)
{
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
if (___update_load_sum(now, &rq->avg_rt,
2018-06-28 15:45:05 +00:00
running,
running,
running)) {
___update_load_avg(&rq->avg_rt, 1, 1);
return 1;
}
return 0;
}
/*
* dl_rq:
*
* util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
* util_sum = cpu_scale * load_sum
* runnable_load_sum = load_sum
*
*/
int update_dl_rq_load_avg(u64 now, struct rq *rq, int running)
{
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
if (___update_load_sum(now, &rq->avg_dl,
running,
running,
running)) {
___update_load_avg(&rq->avg_dl, 1, 1);
return 1;
}
return 0;
}
sched/irq: Add IRQ utilization tracking interrupt and steal time are the only remaining activities tracked by rt_avg. Like for sched classes, we can use PELT to track their average utilization of the CPU. But unlike sched class, we don't track when entering/leaving interrupt; Instead, we take into account the time spent under interrupt context when we update rqs' clock (rq_clock_task). This also means that we have to decay the normal context time and account for interrupt time during the update. That's also important to note that because: rq_clock == rq_clock_task + interrupt time and rq_clock_task is used by a sched class to compute its utilization, the util_avg of a sched class only reflects the utilization of the time spent in normal context and not of the whole time of the CPU. The utilization of interrupt gives an more accurate level of utilization of CPU. The CPU utilization is: avg_irq + (1 - avg_irq / max capacity) * /Sum avg_rq Most of the time, avg_irq is small and neglictible so the use of the approximation CPU utilization = /Sum avg_rq was enough. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: claudio@evidence.eu.com Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: luca.abeni@santannapisa.it Cc: patrick.bellasi@arm.com Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: valentin.schneider@arm.com Cc: viresh.kumar@linaro.org Link: http://lkml.kernel.org/r/1530200714-4504-7-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-28 15:45:09 +00:00
#ifdef CONFIG_HAVE_SCHED_AVG_IRQ
sched/irq: Add IRQ utilization tracking interrupt and steal time are the only remaining activities tracked by rt_avg. Like for sched classes, we can use PELT to track their average utilization of the CPU. But unlike sched class, we don't track when entering/leaving interrupt; Instead, we take into account the time spent under interrupt context when we update rqs' clock (rq_clock_task). This also means that we have to decay the normal context time and account for interrupt time during the update. That's also important to note that because: rq_clock == rq_clock_task + interrupt time and rq_clock_task is used by a sched class to compute its utilization, the util_avg of a sched class only reflects the utilization of the time spent in normal context and not of the whole time of the CPU. The utilization of interrupt gives an more accurate level of utilization of CPU. The CPU utilization is: avg_irq + (1 - avg_irq / max capacity) * /Sum avg_rq Most of the time, avg_irq is small and neglictible so the use of the approximation CPU utilization = /Sum avg_rq was enough. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: claudio@evidence.eu.com Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: luca.abeni@santannapisa.it Cc: patrick.bellasi@arm.com Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: valentin.schneider@arm.com Cc: viresh.kumar@linaro.org Link: http://lkml.kernel.org/r/1530200714-4504-7-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-28 15:45:09 +00:00
/*
* irq:
*
* util_sum = \Sum se->avg.util_sum but se->avg.util_sum is not tracked
* util_sum = cpu_scale * load_sum
* runnable_load_sum = load_sum
*
*/
int update_irq_load_avg(struct rq *rq, u64 running)
{
int ret = 0;
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
/*
* We can't use clock_pelt because irq time is not accounted in
* clock_task. Instead we directly scale the running time to
* reflect the real amount of computation
*/
running = cap_scale(running, arch_scale_freq_capacity(cpu_of(rq)));
running = cap_scale(running, arch_scale_cpu_capacity(NULL, cpu_of(rq)));
sched/irq: Add IRQ utilization tracking interrupt and steal time are the only remaining activities tracked by rt_avg. Like for sched classes, we can use PELT to track their average utilization of the CPU. But unlike sched class, we don't track when entering/leaving interrupt; Instead, we take into account the time spent under interrupt context when we update rqs' clock (rq_clock_task). This also means that we have to decay the normal context time and account for interrupt time during the update. That's also important to note that because: rq_clock == rq_clock_task + interrupt time and rq_clock_task is used by a sched class to compute its utilization, the util_avg of a sched class only reflects the utilization of the time spent in normal context and not of the whole time of the CPU. The utilization of interrupt gives an more accurate level of utilization of CPU. The CPU utilization is: avg_irq + (1 - avg_irq / max capacity) * /Sum avg_rq Most of the time, avg_irq is small and neglictible so the use of the approximation CPU utilization = /Sum avg_rq was enough. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: claudio@evidence.eu.com Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: luca.abeni@santannapisa.it Cc: patrick.bellasi@arm.com Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: valentin.schneider@arm.com Cc: viresh.kumar@linaro.org Link: http://lkml.kernel.org/r/1530200714-4504-7-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-28 15:45:09 +00:00
/*
* We know the time that has been used by interrupt since last update
* but we don't when. Let be pessimistic and assume that interrupt has
* happened just before the update. This is not so far from reality
* because interrupt will most probably wake up task and trig an update
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
* of rq clock during which the metric is updated.
sched/irq: Add IRQ utilization tracking interrupt and steal time are the only remaining activities tracked by rt_avg. Like for sched classes, we can use PELT to track their average utilization of the CPU. But unlike sched class, we don't track when entering/leaving interrupt; Instead, we take into account the time spent under interrupt context when we update rqs' clock (rq_clock_task). This also means that we have to decay the normal context time and account for interrupt time during the update. That's also important to note that because: rq_clock == rq_clock_task + interrupt time and rq_clock_task is used by a sched class to compute its utilization, the util_avg of a sched class only reflects the utilization of the time spent in normal context and not of the whole time of the CPU. The utilization of interrupt gives an more accurate level of utilization of CPU. The CPU utilization is: avg_irq + (1 - avg_irq / max capacity) * /Sum avg_rq Most of the time, avg_irq is small and neglictible so the use of the approximation CPU utilization = /Sum avg_rq was enough. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: claudio@evidence.eu.com Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: luca.abeni@santannapisa.it Cc: patrick.bellasi@arm.com Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: valentin.schneider@arm.com Cc: viresh.kumar@linaro.org Link: http://lkml.kernel.org/r/1530200714-4504-7-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-28 15:45:09 +00:00
* We start to decay with normal context time and then we add the
* interrupt context time.
* We can safely remove running from rq->clock because
* rq->clock += delta with delta >= running
*/
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
ret = ___update_load_sum(rq->clock - running, &rq->avg_irq,
sched/irq: Add IRQ utilization tracking interrupt and steal time are the only remaining activities tracked by rt_avg. Like for sched classes, we can use PELT to track their average utilization of the CPU. But unlike sched class, we don't track when entering/leaving interrupt; Instead, we take into account the time spent under interrupt context when we update rqs' clock (rq_clock_task). This also means that we have to decay the normal context time and account for interrupt time during the update. That's also important to note that because: rq_clock == rq_clock_task + interrupt time and rq_clock_task is used by a sched class to compute its utilization, the util_avg of a sched class only reflects the utilization of the time spent in normal context and not of the whole time of the CPU. The utilization of interrupt gives an more accurate level of utilization of CPU. The CPU utilization is: avg_irq + (1 - avg_irq / max capacity) * /Sum avg_rq Most of the time, avg_irq is small and neglictible so the use of the approximation CPU utilization = /Sum avg_rq was enough. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: claudio@evidence.eu.com Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: luca.abeni@santannapisa.it Cc: patrick.bellasi@arm.com Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: valentin.schneider@arm.com Cc: viresh.kumar@linaro.org Link: http://lkml.kernel.org/r/1530200714-4504-7-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-28 15:45:09 +00:00
0,
0,
0);
sched/fair: Update scale invariance of PELT The current implementation of load tracking invariance scales the contribution with current frequency and uarch performance (only for utilization) of the CPU. One main result of this formula is that the figures are capped by current capacity of CPU. Another one is that the load_avg is not invariant because not scaled with uarch. The util_avg of a periodic task that runs r time slots every p time slots varies in the range : U * (1-y^r)/(1-y^p) * y^i < Utilization < U * (1-y^r)/(1-y^p) with U is the max util_avg value = SCHED_CAPACITY_SCALE At a lower capacity, the range becomes: U * C * (1-y^r')/(1-y^p) * y^i' < Utilization < U * C * (1-y^r')/(1-y^p) with C reflecting the compute capacity ratio between current capacity and max capacity. so C tries to compensate changes in (1-y^r') but it can't be accurate. Instead of scaling the contribution value of PELT algo, we should scale the running time. The PELT signal aims to track the amount of computation of tasks and/or rq so it seems more correct to scale the running time to reflect the effective amount of computation done since the last update. In order to be fully invariant, we need to apply the same amount of running time and idle time whatever the current capacity. Because running at lower capacity implies that the task will run longer, we have to ensure that the same amount of idle time will be applied when system becomes idle and no idle time has been "stolen". But reaching the maximum utilization value (SCHED_CAPACITY_SCALE) means that the task is seen as an always-running task whatever the capacity of the CPU (even at max compute capacity). In this case, we can discard this "stolen" idle times which becomes meaningless. In order to achieve this time scaling, a new clock_pelt is created per rq. The increase of this clock scales with current capacity when something is running on rq and synchronizes with clock_task when rq is idle. With this mechanism, we ensure the same running and idle time whatever the current capacity. This also enables to simplify the pelt algorithm by removing all references of uarch and frequency and applying the same contribution to utilization and loads. Furthermore, the scaling is done only once per update of clock (update_rq_clock_task()) instead of during each update of sched_entities and cfs/rt/dl_rq of the rq like the current implementation. This is interesting when cgroup are involved as shown in the results below: On a hikey (octo Arm64 platform). Performance cpufreq governor and only shallowest c-state to remove variance generated by those power features so we only track the impact of pelt algo. each test runs 16 times: ./perf bench sched pipe (higher is better) kernel tip/sched/core + patch ops/seconds ops/seconds diff cgroup root 59652(+/- 0.18%) 59876(+/- 0.24%) +0.38% level1 55608(+/- 0.27%) 55923(+/- 0.24%) +0.57% level2 52115(+/- 0.29%) 52564(+/- 0.22%) +0.86% hackbench -l 1000 (lower is better) kernel tip/sched/core + patch duration(sec) duration(sec) diff cgroup root 4.453(+/- 2.37%) 4.383(+/- 2.88%) -1.57% level1 4.859(+/- 8.50%) 4.830(+/- 7.07%) -0.60% level2 5.063(+/- 9.83%) 4.928(+/- 9.66%) -2.66% Then, the responsiveness of PELT is improved when CPU is not running at max capacity with this new algorithm. I have put below some examples of duration to reach some typical load values according to the capacity of the CPU with current implementation and with this patch. These values has been computed based on the geometric series and the half period value: Util (%) max capacity half capacity(mainline) half capacity(w/ patch) 972 (95%) 138ms not reachable 276ms 486 (47.5%) 30ms 138ms 60ms 256 (25%) 13ms 32ms 26ms On my hikey (octo Arm64 platform) with schedutil governor, the time to reach max OPP when starting from a null utilization, decreases from 223ms with current scale invariance down to 121ms with the new algorithm. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: patrick.bellasi@arm.com Cc: pjt@google.com Cc: pkondeti@codeaurora.org Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: srinivas.pandruvada@linux.intel.com Cc: thara.gopinath@linaro.org Link: https://lkml.kernel.org/r/1548257214-13745-3-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2019-01-23 15:26:53 +00:00
ret += ___update_load_sum(rq->clock, &rq->avg_irq,
sched/irq: Add IRQ utilization tracking interrupt and steal time are the only remaining activities tracked by rt_avg. Like for sched classes, we can use PELT to track their average utilization of the CPU. But unlike sched class, we don't track when entering/leaving interrupt; Instead, we take into account the time spent under interrupt context when we update rqs' clock (rq_clock_task). This also means that we have to decay the normal context time and account for interrupt time during the update. That's also important to note that because: rq_clock == rq_clock_task + interrupt time and rq_clock_task is used by a sched class to compute its utilization, the util_avg of a sched class only reflects the utilization of the time spent in normal context and not of the whole time of the CPU. The utilization of interrupt gives an more accurate level of utilization of CPU. The CPU utilization is: avg_irq + (1 - avg_irq / max capacity) * /Sum avg_rq Most of the time, avg_irq is small and neglictible so the use of the approximation CPU utilization = /Sum avg_rq was enough. Signed-off-by: Vincent Guittot <vincent.guittot@linaro.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten.Rasmussen@arm.com Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: claudio@evidence.eu.com Cc: daniel.lezcano@linaro.org Cc: dietmar.eggemann@arm.com Cc: joel@joelfernandes.org Cc: juri.lelli@redhat.com Cc: luca.abeni@santannapisa.it Cc: patrick.bellasi@arm.com Cc: quentin.perret@arm.com Cc: rjw@rjwysocki.net Cc: valentin.schneider@arm.com Cc: viresh.kumar@linaro.org Link: http://lkml.kernel.org/r/1530200714-4504-7-git-send-email-vincent.guittot@linaro.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-06-28 15:45:09 +00:00
1,
1,
1);
if (ret)
___update_load_avg(&rq->avg_irq, 1, 1);
return ret;
}
#endif