Based on kernel version 4.16.1. Page generated on 2018-04-09 11:53 EST.
1 Power allocator governor tunables 2 ================================= 3 4 Trip points 5 ----------- 6 7 The governor works optimally with the following two passive trip points: 8 9 1. "switch on" trip point: temperature above which the governor 10 control loop starts operating. This is the first passive trip 11 point of the thermal zone. 12 13 2. "desired temperature" trip point: it should be higher than the 14 "switch on" trip point. This the target temperature the governor 15 is controlling for. This is the last passive trip point of the 16 thermal zone. 17 18 PID Controller 19 -------------- 20 21 The power allocator governor implements a 22 Proportional-Integral-Derivative controller (PID controller) with 23 temperature as the control input and power as the controlled output: 24 25 P_max = k_p * e + k_i * err_integral + k_d * diff_err + sustainable_power 26 27 where 28 e = desired_temperature - current_temperature 29 err_integral is the sum of previous errors 30 diff_err = e - previous_error 31 32 It is similar to the one depicted below: 33 34 k_d 35 | 36 current_temp | 37 | v 38 | +----------+ +---+ 39 | +----->| diff_err |-->| X |------+ 40 | | +----------+ +---+ | 41 | | | tdp actor 42 | | k_i | | get_requested_power() 43 | | | | | | | 44 | | | | | | | ... 45 v | v v v v v 46 +---+ | +-------+ +---+ +---+ +---+ +----------+ 47 | S |-------+----->| sum e |----->| X |--->| S |-->| S |-->|power | 48 +---+ | +-------+ +---+ +---+ +---+ |allocation| 49 ^ | ^ +----------+ 50 | | | | | 51 | | +---+ | | | 52 | +------->| X |-------------------+ v v 53 | +---+ granted performance 54 desired_temperature ^ 55 | 56 | 57 k_po/k_pu 58 59 Sustainable power 60 ----------------- 61 62 An estimate of the sustainable dissipatable power (in mW) should be 63 provided while registering the thermal zone. This estimates the 64 sustained power that can be dissipated at the desired control 65 temperature. This is the maximum sustained power for allocation at 66 the desired maximum temperature. The actual sustained power can vary 67 for a number of reasons. The closed loop controller will take care of 68 variations such as environmental conditions, and some factors related 69 to the speed-grade of the silicon. `sustainable_power` is therefore 70 simply an estimate, and may be tuned to affect the aggressiveness of 71 the thermal ramp. For reference, the sustainable power of a 4" phone 72 is typically 2000mW, while on a 10" tablet is around 4500mW (may vary 73 depending on screen size). 74 75 If you are using device tree, do add it as a property of the 76 thermal-zone. For example: 77 78 thermal-zones { 79 soc_thermal { 80 polling-delay = <1000>; 81 polling-delay-passive = <100>; 82 sustainable-power = <2500>; 83 ... 84 85 Instead, if the thermal zone is registered from the platform code, pass a 86 `thermal_zone_params` that has a `sustainable_power`. If no 87 `thermal_zone_params` were being passed, then something like below 88 will suffice: 89 90 static const struct thermal_zone_params tz_params = { 91 .sustainable_power = 3500, 92 }; 93 94 and then pass `tz_params` as the 5th parameter to 95 `thermal_zone_device_register()` 96 97 k_po and k_pu 98 ------------- 99 100 The implementation of the PID controller in the power allocator 101 thermal governor allows the configuration of two proportional term 102 constants: `k_po` and `k_pu`. `k_po` is the proportional term 103 constant during temperature overshoot periods (current temperature is 104 above "desired temperature" trip point). Conversely, `k_pu` is the 105 proportional term constant during temperature undershoot periods 106 (current temperature below "desired temperature" trip point). 107 108 These controls are intended as the primary mechanism for configuring 109 the permitted thermal "ramp" of the system. For instance, a lower 110 `k_pu` value will provide a slower ramp, at the cost of capping 111 available capacity at a low temperature. On the other hand, a high 112 value of `k_pu` will result in the governor granting very high power 113 whilst temperature is low, and may lead to temperature overshooting. 114 115 The default value for `k_pu` is: 116 117 2 * sustainable_power / (desired_temperature - switch_on_temp) 118 119 This means that at `switch_on_temp` the output of the controller's 120 proportional term will be 2 * `sustainable_power`. The default value 121 for `k_po` is: 122 123 sustainable_power / (desired_temperature - switch_on_temp) 124 125 Focusing on the proportional and feed forward values of the PID 126 controller equation we have: 127 128 P_max = k_p * e + sustainable_power 129 130 The proportional term is proportional to the difference between the 131 desired temperature and the current one. When the current temperature 132 is the desired one, then the proportional component is zero and 133 `P_max` = `sustainable_power`. That is, the system should operate in 134 thermal equilibrium under constant load. `sustainable_power` is only 135 an estimate, which is the reason for closed-loop control such as this. 136 137 Expanding `k_pu` we get: 138 P_max = 2 * sustainable_power * (T_set - T) / (T_set - T_on) + 139 sustainable_power 140 141 where 142 T_set is the desired temperature 143 T is the current temperature 144 T_on is the switch on temperature 145 146 When the current temperature is the switch_on temperature, the above 147 formula becomes: 148 149 P_max = 2 * sustainable_power * (T_set - T_on) / (T_set - T_on) + 150 sustainable_power = 2 * sustainable_power + sustainable_power = 151 3 * sustainable_power 152 153 Therefore, the proportional term alone linearly decreases power from 154 3 * `sustainable_power` to `sustainable_power` as the temperature 155 rises from the switch on temperature to the desired temperature. 156 157 k_i and integral_cutoff 158 ----------------------- 159 160 `k_i` configures the PID loop's integral term constant. This term 161 allows the PID controller to compensate for long term drift and for 162 the quantized nature of the output control: cooling devices can't set 163 the exact power that the governor requests. When the temperature 164 error is below `integral_cutoff`, errors are accumulated in the 165 integral term. This term is then multiplied by `k_i` and the result 166 added to the output of the controller. Typically `k_i` is set low (1 167 or 2) and `integral_cutoff` is 0. 168 169 k_d 170 --- 171 172 `k_d` configures the PID loop's derivative term constant. It's 173 recommended to leave it as the default: 0. 174 175 Cooling device power API 176 ======================== 177 178 Cooling devices controlled by this governor must supply the additional 179 "power" API in their `cooling_device_ops`. It consists on three ops: 180 181 1. int get_requested_power(struct thermal_cooling_device *cdev, 182 struct thermal_zone_device *tz, u32 *power); 183 @cdev: The `struct thermal_cooling_device` pointer 184 @tz: thermal zone in which we are currently operating 185 @power: pointer in which to store the calculated power 186 187 `get_requested_power()` calculates the power requested by the device 188 in milliwatts and stores it in @power . It should return 0 on 189 success, -E* on failure. This is currently used by the power 190 allocator governor to calculate how much power to give to each cooling 191 device. 192 193 2. int state2power(struct thermal_cooling_device *cdev, struct 194 thermal_zone_device *tz, unsigned long state, u32 *power); 195 @cdev: The `struct thermal_cooling_device` pointer 196 @tz: thermal zone in which we are currently operating 197 @state: A cooling device state 198 @power: pointer in which to store the equivalent power 199 200 Convert cooling device state @state into power consumption in 201 milliwatts and store it in @power. It should return 0 on success, -E* 202 on failure. This is currently used by thermal core to calculate the 203 maximum power that an actor can consume. 204 205 3. int power2state(struct thermal_cooling_device *cdev, u32 power, 206 unsigned long *state); 207 @cdev: The `struct thermal_cooling_device` pointer 208 @power: power in milliwatts 209 @state: pointer in which to store the resulting state 210 211 Calculate a cooling device state that would make the device consume at 212 most @power mW and store it in @state. It should return 0 on success, 213 -E* on failure. This is currently used by the thermal core to convert 214 a given power set by the power allocator governor to a state that the 215 cooling device can set. It is a function because this conversion may 216 depend on external factors that may change so this function should the 217 best conversion given "current circumstances". 218 219 Cooling device weights 220 ---------------------- 221 222 Weights are a mechanism to bias the allocation among cooling 223 devices. They express the relative power efficiency of different 224 cooling devices. Higher weight can be used to express higher power 225 efficiency. Weighting is relative such that if each cooling device 226 has a weight of one they are considered equal. This is particularly 227 useful in heterogeneous systems where two cooling devices may perform 228 the same kind of compute, but with different efficiency. For example, 229 a system with two different types of processors. 230 231 If the thermal zone is registered using 232 `thermal_zone_device_register()` (i.e., platform code), then weights 233 are passed as part of the thermal zone's `thermal_bind_parameters`. 234 If the platform is registered using device tree, then they are passed 235 as the `contribution` property of each map in the `cooling-maps` node. 236 237 Limitations of the power allocator governor 238 =========================================== 239 240 The power allocator governor's PID controller works best if there is a 241 periodic tick. If you have a driver that calls 242 `thermal_zone_device_update()` (or anything that ends up calling the 243 governor's `throttle()` function) repetitively, the governor response 244 won't be very good. Note that this is not particular to this 245 governor, step-wise will also misbehave if you call its throttle() 246 faster than the normal thermal framework tick (due to interrupts for 247 example) as it will overreact.