1 =========
2 Schedutil
3 =========
4
5 .. note::
6
7 All this assumes a linear relation between frequency and work capacity,
8 we know this is flawed, but it is the best workable approximation.
9
10
11 PELT (Per Entity Load Tracking)
12 ===============================
13
14 With PELT we track some metrics across the various scheduler entities, from
15 individual tasks to task-group slices to CPU runqueues. As the basis for this
16 we use an Exponentially Weighted Moving Average (EWMA), each period (1024us)
17 is decayed such that y^32 = 0.5. That is, the most recent 32ms contribute
18 half, while the rest of history contribute the other half.
19
20 Specifically:
21
22 ewma_sum(u) := u_0 + u_1*y + u_2*y^2 + ...
23
24 ewma(u) = ewma_sum(u) / ewma_sum(1)
25
26 Since this is essentially a progression of an infinite geometric series, the
27 results are composable, that is ewma(A) + ewma(B) = ewma(A+B). This property
28 is key, since it gives the ability to recompose the averages when tasks move
29 around.
30
31 Note that blocked tasks still contribute to the aggregates (task-group slices
32 and CPU runqueues), which reflects their expected contribution when they
33 resume running.
34
35 Using this we track 2 key metrics: 'running' and 'runnable'. 'Running'
36 reflects the time an entity spends on the CPU, while 'runnable' reflects the
37 time an entity spends on the runqueue. When there is only a single task these
38 two metrics are the same, but once there is contention for the CPU 'running'
39 will decrease to reflect the fraction of time each task spends on the CPU
40 while 'runnable' will increase to reflect the amount of contention.
41
42 For more detail see: kernel/sched/pelt.c
43
44
45 Frequency / CPU Invariance
46 ==========================
47
48 Because consuming the CPU for 50% at 1GHz is not the same as consuming the CPU
49 for 50% at 2GHz, nor is running 50% on a LITTLE CPU the same as running 50% on
50 a big CPU, we allow architectures to scale the time delta with two ratios, one
51 Dynamic Voltage and Frequency Scaling (DVFS) ratio and one microarch ratio.
52
53 For simple DVFS architectures (where software is in full control) we trivially
54 compute the ratio as::
55
56 f_cur
57 r_dvfs := -----
58 f_max
59
60 For more dynamic systems where the hardware is in control of DVFS we use
61 hardware counters (Intel APERF/MPERF, ARMv8.4-AMU) to provide us this ratio.
62 For Intel specifically, we use::
63
64 APERF
65 f_cur := ----- * P0
66 MPERF
67
68 4C-turbo; if available and turbo enabled
69 f_max := { 1C-turbo; if turbo enabled
70 P0; otherwise
71
72 f_cur
73 r_dvfs := min( 1, ----- )
74 f_max
75
76 We pick 4C turbo over 1C turbo to make it slightly more sustainable.
77
78 r_cpu is determined as the ratio of highest performance level of the current
79 CPU vs the highest performance level of any other CPU in the system.
80
81 r_tot = r_dvfs * r_cpu
82
83 The result is that the above 'running' and 'runnable' metrics become invariant
84 of DVFS and CPU type. IOW. we can transfer and compare them between CPUs.
85
86 For more detail see:
87
88 - kernel/sched/pelt.h:update_rq_clock_pelt()
89 - arch/x86/kernel/smpboot.c:"APERF/MPERF frequency ratio computation."
90 - Documentation/scheduler/sched-capacity.rst:"1. CPU Capacity + 2. Task utilization"
91
92
93 UTIL_EST
94 ========
95
96 Because periodic tasks have their averages decayed while they sleep, even
97 though when running their expected utilization will be the same, they suffer a
98 (DVFS) ramp-up after they are running again.
99
100 To alleviate this (a default enabled option) UTIL_EST drives an Infinite
101 Impulse Response (IIR) EWMA with the 'running' value on dequeue -- when it is
102 highest. UTIL_EST filters to instantly increase and only decay on decrease.
103
104 A further runqueue wide sum (of runnable tasks) is maintained of:
105
106 util_est := \Sum_t max( t_running, t_util_est_ewma )
107
108 For more detail see: kernel/sched/fair.c:util_est_dequeue()
109
110
111 UCLAMP
112 ======
113
114 It is possible to set effective u_min and u_max clamps on each CFS or RT task;
115 the runqueue keeps an max aggregate of these clamps for all running tasks.
116
117 For more detail see: include/uapi/linux/sched/types.h
118
119
120 Schedutil / DVFS
121 ================
122
123 Every time the scheduler load tracking is updated (task wakeup, task
124 migration, time progression) we call out to schedutil to update the hardware
125 DVFS state.
126
127 The basis is the CPU runqueue's 'running' metric, which per the above it is
128 the frequency invariant utilization estimate of the CPU. From this we compute
129 a desired frequency like::
130
131 max( running, util_est ); if UTIL_EST
132 u_cfs := { running; otherwise
133
134 clamp( u_cfs + u_rt , u_min, u_max ); if UCLAMP_TASK
135 u_clamp := { u_cfs + u_rt; otherwise
136
137 u := u_clamp + u_irq + u_dl; [approx. see source for more detail]
138
139 f_des := min( f_max, 1.25 u * f_max )
140
141 XXX IO-wait: when the update is due to a task wakeup from IO-completion we
142 boost 'u' above.
143
144 This frequency is then used to select a P-state/OPP or directly munged into a
145 CPPC style request to the hardware.
146
147 XXX: deadline tasks (Sporadic Task Model) allows us to calculate a hard f_min
148 required to satisfy the workload.
149
150 Because these callbacks are directly from the scheduler, the DVFS hardware
151 interaction should be 'fast' and non-blocking. Schedutil supports
152 rate-limiting DVFS requests for when hardware interaction is slow and
153 expensive, this reduces effectiveness.
154
155 For more information see: kernel/sched/cpufreq_schedutil.c
156
157
158 NOTES
159 =====
160
161 - On low-load scenarios, where DVFS is most relevant, the 'running' numbers
162 will closely reflect utilization.
163
164 - In saturated scenarios task movement will cause some transient dips,
165 suppose we have a CPU saturated with 4 tasks, then when we migrate a task
166 to an idle CPU, the old CPU will have a 'running' value of 0.75 while the
167 new CPU will gain 0.25. This is inevitable and time progression will
168 correct this. XXX do we still guarantee f_max due to no idle-time?
169
170 - Much of the above is about avoiding DVFS dips, and independent DVFS domains
171 having to re-learn / ramp-up when load shifts.
172
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.