| 1 |
.. SPDX-License-Identifier: GPL-2.0 |
| 2 |
|
| 3 |
================ |
| 4 |
CPU Idle Cooling |
| 5 |
================ |
| 6 |
|
| 7 |
Situation: |
| 8 |
---------- |
| 9 |
|
| 10 |
Under certain circumstances a SoC can reach a critical temperature |
| 11 |
limit and is unable to stabilize the temperature around a temperature |
| 12 |
control. When the SoC has to stabilize the temperature, the kernel can |
| 13 |
act on a cooling device to mitigate the dissipated power. When the |
| 14 |
critical temperature is reached, a decision must be taken to reduce |
| 15 |
the temperature, that, in turn impacts performance. |
| 16 |
|
| 17 |
Another situation is when the silicon temperature continues to |
| 18 |
increase even after the dynamic leakage is reduced to its minimum by |
| 19 |
clock gating the component. This runaway phenomenon can continue due |
| 20 |
to the static leakage. The only solution is to power down the |
| 21 |
component, thus dropping the dynamic and static leakage that will |
| 22 |
allow the component to cool down. |
| 23 |
|
| 24 |
Last but not least, the system can ask for a specific power budget but |
| 25 |
because of the OPP density, we can only choose an OPP with a power |
| 26 |
budget lower than the requested one and under-utilize the CPU, thus |
| 27 |
losing performance. In other words, one OPP under-utilizes the CPU |
| 28 |
with a power less than the requested power budget and the next OPP |
| 29 |
exceeds the power budget. An intermediate OPP could have been used if |
| 30 |
it were present. |
| 31 |
|
| 32 |
Solutions: |
| 33 |
---------- |
| 34 |
|
| 35 |
If we can remove the static and the dynamic leakage for a specific |
| 36 |
duration in a controlled period, the SoC temperature will |
| 37 |
decrease. Acting on the idle state duration or the idle cycle |
| 38 |
injection period, we can mitigate the temperature by modulating the |
| 39 |
power budget. |
| 40 |
|
| 41 |
The Operating Performance Point (OPP) density has a great influence on |
| 42 |
the control precision of cpufreq, however different vendors have a |
| 43 |
plethora of OPP density, and some have large power gap between OPPs, |
| 44 |
that will result in loss of performance during thermal control and |
| 45 |
loss of power in other scenarios. |
| 46 |
|
| 47 |
At a specific OPP, we can assume that injecting idle cycle on all CPUs |
| 48 |
belong to the same cluster, with a duration greater than the cluster |
| 49 |
idle state target residency, we lead to dropping the static and the |
| 50 |
dynamic leakage for this period (modulo the energy needed to enter |
| 51 |
this state). So the sustainable power with idle cycles has a linear |
| 52 |
relation with the OPP’s sustainable power and can be computed with a |
| 53 |
coefficient similar to:: |
| 54 |
|
| 55 |
Power(IdleCycle) = Coef x Power(OPP) |
| 56 |
|
| 57 |
Idle Injection: |
| 58 |
--------------- |
| 59 |
|
| 60 |
The base concept of the idle injection is to force the CPU to go to an |
| 61 |
idle state for a specified time each control cycle, it provides |
| 62 |
another way to control CPU power and heat in addition to |
| 63 |
cpufreq. Ideally, if all CPUs belonging to the same cluster, inject |
| 64 |
their idle cycles synchronously, the cluster can reach its power down |
| 65 |
state with a minimum power consumption and reduce the static leakage |
| 66 |
to almost zero. However, these idle cycles injection will add extra |
| 67 |
latencies as the CPUs will have to wakeup from a deep sleep state. |
| 68 |
|
| 69 |
We use a fixed duration of idle injection that gives an acceptable |
| 70 |
performance penalty and a fixed latency. Mitigation can be increased |
| 71 |
or decreased by modulating the duty cycle of the idle injection. |
| 72 |
|
| 73 |
:: |
| 74 |
|
| 75 |
^ |
| 76 |
| |
| 77 |
| |
| 78 |
|------- ------- |
| 79 |
|_______|_______________________|_______|___________ |
| 80 |
|
| 81 |
<------> |
| 82 |
idle <----------------------> |
| 83 |
running |
| 84 |
|
| 85 |
<-----------------------------> |
| 86 |
duty cycle 25% |
| 87 |
|
| 88 |
|
| 89 |
The implementation of the cooling device bases the number of states on |
| 90 |
the duty cycle percentage. When no mitigation is happening the cooling |
| 91 |
device state is zero, meaning the duty cycle is 0%. |
| 92 |
|
| 93 |
When the mitigation begins, depending on the governor's policy, a |
| 94 |
starting state is selected. With a fixed idle duration and the duty |
| 95 |
cycle (aka the cooling device state), the running duration can be |
| 96 |
computed. |
| 97 |
|
| 98 |
The governor will change the cooling device state thus the duty cycle |
| 99 |
and this variation will modulate the cooling effect. |
| 100 |
|
| 101 |
:: |
| 102 |
|
| 103 |
^ |
| 104 |
| |
| 105 |
| |
| 106 |
|------- ------- |
| 107 |
|_______|_______________|_______|___________ |
| 108 |
|
| 109 |
<------> |
| 110 |
idle <--------------> |
| 111 |
running |
| 112 |
|
| 113 |
<---------------------> |
| 114 |
duty cycle 33% |
| 115 |
|
| 116 |
|
| 117 |
^ |
| 118 |
| |
| 119 |
| |
| 120 |
|------- ------- |
| 121 |
|_______|_______|_______|___________ |
| 122 |
|
| 123 |
<------> |
| 124 |
idle <------> |
| 125 |
running |
| 126 |
|
| 127 |
<-------------> |
| 128 |
duty cycle 50% |
| 129 |
|
| 130 |
The idle injection duration value must comply with the constraints: |
| 131 |
|
| 132 |
- It is less than or equal to the latency we tolerate when the |
| 133 |
mitigation begins. It is platform dependent and will depend on the |
| 134 |
user experience, reactivity vs performance trade off we want. This |
| 135 |
value should be specified. |
| 136 |
|
| 137 |
- It is greater than the idle state’s target residency we want to go |
| 138 |
for thermal mitigation, otherwise we end up consuming more energy. |
| 139 |
|
| 140 |
Power considerations |
| 141 |
-------------------- |
| 142 |
|
| 143 |
When we reach the thermal trip point, we have to sustain a specified |
| 144 |
power for a specific temperature but at this time we consume:: |
| 145 |
|
| 146 |
Power = Capacitance x Voltage^2 x Frequency x Utilisation |
| 147 |
|
| 148 |
... which is more than the sustainable power (or there is something |
| 149 |
wrong in the system setup). The ‘Capacitance’ and ‘Utilisation’ are a |
| 150 |
fixed value, ‘Voltage’ and the ‘Frequency’ are fixed artificially |
| 151 |
because we don’t want to change the OPP. We can group the |
| 152 |
‘Capacitance’ and the ‘Utilisation’ into a single term which is the |
| 153 |
‘Dynamic Power Coefficient (Cdyn)’ Simplifying the above, we have:: |
| 154 |
|
| 155 |
Pdyn = Cdyn x Voltage^2 x Frequency |
| 156 |
|
| 157 |
The power allocator governor will ask us somehow to reduce our power |
| 158 |
in order to target the sustainable power defined in the device |
| 159 |
tree. So with the idle injection mechanism, we want an average power |
| 160 |
(Ptarget) resulting in an amount of time running at full power on a |
| 161 |
specific OPP and idle another amount of time. That could be put in a |
| 162 |
equation:: |
| 163 |
|
| 164 |
P(opp)target = ((Trunning x (P(opp)running) + (Tidle x P(opp)idle)) / |
| 165 |
(Trunning + Tidle) |
| 166 |
|
| 167 |
... |
| 168 |
|
| 169 |
Tidle = Trunning x ((P(opp)running / P(opp)target) - 1) |
| 170 |
|
| 171 |
At this point if we know the running period for the CPU, that gives us |
| 172 |
the idle injection we need. Alternatively if we have the idle |
| 173 |
injection duration, we can compute the running duration with:: |
| 174 |
|
| 175 |
Trunning = Tidle / ((P(opp)running / P(opp)target) - 1) |
| 176 |
|
| 177 |
Practically, if the running power is less than the targeted power, we |
| 178 |
end up with a negative time value, so obviously the equation usage is |
| 179 |
bound to a power reduction, hence a higher OPP is needed to have the |
| 180 |
running power greater than the targeted power. |
| 181 |
|
| 182 |
However, in this demonstration we ignore three aspects: |
| 183 |
|
| 184 |
* The static leakage is not defined here, we can introduce it in the |
| 185 |
equation but assuming it will be zero most of the time as it is |
| 186 |
difficult to get the values from the SoC vendors |
| 187 |
|
| 188 |
* The idle state wake up latency (or entry + exit latency) is not |
| 189 |
taken into account, it must be added in the equation in order to |
| 190 |
rigorously compute the idle injection |
| 191 |
|
| 192 |
* The injected idle duration must be greater than the idle state |
| 193 |
target residency, otherwise we end up consuming more energy and |
| 194 |
potentially invert the mitigation effect |
| 195 |
|
| 196 |
So the final equation is:: |
| 197 |
|
| 198 |
Trunning = (Tidle - Twakeup ) x |
| 199 |
(((P(opp)dyn + P(opp)static ) - P(opp)target) / P(opp)target ) |