| 1 |
======================= |
| 2 |
Intel Powerclamp Driver |
| 3 |
======================= |
| 4 |
|
| 5 |
By: |
| 6 |
- Arjan van de Ven <[email protected]> |
| 7 |
- Jacob Pan <[email protected]> |
| 8 |
|
| 9 |
.. Contents: |
| 10 |
|
| 11 |
(*) Introduction |
| 12 |
- Goals and Objectives |
| 13 |
|
| 14 |
(*) Theory of Operation |
| 15 |
- Idle Injection |
| 16 |
- Calibration |
| 17 |
|
| 18 |
(*) Performance Analysis |
| 19 |
- Effectiveness and Limitations |
| 20 |
- Power vs Performance |
| 21 |
- Scalability |
| 22 |
- Calibration |
| 23 |
- Comparison with Alternative Techniques |
| 24 |
|
| 25 |
(*) Usage and Interfaces |
| 26 |
- Generic Thermal Layer (sysfs) |
| 27 |
- Kernel APIs (TBD) |
| 28 |
|
| 29 |
(*) Module Parameters |
| 30 |
|
| 31 |
INTRODUCTION |
| 32 |
============ |
| 33 |
|
| 34 |
Consider the situation where a system’s power consumption must be |
| 35 |
reduced at runtime, due to power budget, thermal constraint, or noise |
| 36 |
level, and where active cooling is not preferred. Software managed |
| 37 |
passive power reduction must be performed to prevent the hardware |
| 38 |
actions that are designed for catastrophic scenarios. |
| 39 |
|
| 40 |
Currently, P-states, T-states (clock modulation), and CPU offlining |
| 41 |
are used for CPU throttling. |
| 42 |
|
| 43 |
On Intel CPUs, C-states provide effective power reduction, but so far |
| 44 |
they’re only used opportunistically, based on workload. With the |
| 45 |
development of intel_powerclamp driver, the method of synchronizing |
| 46 |
idle injection across all online CPU threads was introduced. The goal |
| 47 |
is to achieve forced and controllable C-state residency. |
| 48 |
|
| 49 |
Test/Analysis has been made in the areas of power, performance, |
| 50 |
scalability, and user experience. In many cases, clear advantage is |
| 51 |
shown over taking the CPU offline or modulating the CPU clock. |
| 52 |
|
| 53 |
|
| 54 |
THEORY OF OPERATION |
| 55 |
=================== |
| 56 |
|
| 57 |
Idle Injection |
| 58 |
-------------- |
| 59 |
|
| 60 |
On modern Intel processors (Nehalem or later), package level C-state |
| 61 |
residency is available in MSRs, thus also available to the kernel. |
| 62 |
|
| 63 |
These MSRs are:: |
| 64 |
|
| 65 |
#define MSR_PKG_C2_RESIDENCY 0x60D |
| 66 |
#define MSR_PKG_C3_RESIDENCY 0x3F8 |
| 67 |
#define MSR_PKG_C6_RESIDENCY 0x3F9 |
| 68 |
#define MSR_PKG_C7_RESIDENCY 0x3FA |
| 69 |
|
| 70 |
If the kernel can also inject idle time to the system, then a |
| 71 |
closed-loop control system can be established that manages package |
| 72 |
level C-state. The intel_powerclamp driver is conceived as such a |
| 73 |
control system, where the target set point is a user-selected idle |
| 74 |
ratio (based on power reduction), and the error is the difference |
| 75 |
between the actual package level C-state residency ratio and the target idle |
| 76 |
ratio. |
| 77 |
|
| 78 |
Injection is controlled by high priority kernel threads, spawned for |
| 79 |
each online CPU. |
| 80 |
|
| 81 |
These kernel threads, with SCHED_FIFO class, are created to perform |
| 82 |
clamping actions of controlled duty ratio and duration. Each per-CPU |
| 83 |
thread synchronizes its idle time and duration, based on the rounding |
| 84 |
of jiffies, so accumulated errors can be prevented to avoid a jittery |
| 85 |
effect. Threads are also bound to the CPU such that they cannot be |
| 86 |
migrated, unless the CPU is taken offline. In this case, threads |
| 87 |
belong to the offlined CPUs will be terminated immediately. |
| 88 |
|
| 89 |
Running as SCHED_FIFO and relatively high priority, also allows such |
| 90 |
scheme to work for both preemptible and non-preemptible kernels. |
| 91 |
Alignment of idle time around jiffies ensures scalability for HZ |
| 92 |
values. This effect can be better visualized using a Perf timechart. |
| 93 |
The following diagram shows the behavior of kernel thread |
| 94 |
kidle_inject/cpu. During idle injection, it runs monitor/mwait idle |
| 95 |
for a given "duration", then relinquishes the CPU to other tasks, |
| 96 |
until the next time interval. |
| 97 |
|
| 98 |
The NOHZ schedule tick is disabled during idle time, but interrupts |
| 99 |
are not masked. Tests show that the extra wakeups from scheduler tick |
| 100 |
have a dramatic impact on the effectiveness of the powerclamp driver |
| 101 |
on large scale systems (Westmere system with 80 processors). |
| 102 |
|
| 103 |
:: |
| 104 |
|
| 105 |
CPU0 |
| 106 |
____________ ____________ |
| 107 |
kidle_inject/0 | sleep | mwait | sleep | |
| 108 |
_________| |________| |_______ |
| 109 |
duration |
| 110 |
CPU1 |
| 111 |
____________ ____________ |
| 112 |
kidle_inject/1 | sleep | mwait | sleep | |
| 113 |
_________| |________| |_______ |
| 114 |
^ |
| 115 |
| |
| 116 |
| |
| 117 |
roundup(jiffies, interval) |
| 118 |
|
| 119 |
Only one CPU is allowed to collect statistics and update global |
| 120 |
control parameters. This CPU is referred to as the controlling CPU in |
| 121 |
this document. The controlling CPU is elected at runtime, with a |
| 122 |
policy that favors BSP, taking into account the possibility of a CPU |
| 123 |
hot-plug. |
| 124 |
|
| 125 |
In terms of dynamics of the idle control system, package level idle |
| 126 |
time is considered largely as a non-causal system where its behavior |
| 127 |
cannot be based on the past or current input. Therefore, the |
| 128 |
intel_powerclamp driver attempts to enforce the desired idle time |
| 129 |
instantly as given input (target idle ratio). After injection, |
| 130 |
powerclamp monitors the actual idle for a given time window and adjust |
| 131 |
the next injection accordingly to avoid over/under correction. |
| 132 |
|
| 133 |
When used in a causal control system, such as a temperature control, |
| 134 |
it is up to the user of this driver to implement algorithms where |
| 135 |
past samples and outputs are included in the feedback. For example, a |
| 136 |
PID-based thermal controller can use the powerclamp driver to |
| 137 |
maintain a desired target temperature, based on integral and |
| 138 |
derivative gains of the past samples. |
| 139 |
|
| 140 |
|
| 141 |
|
| 142 |
Calibration |
| 143 |
----------- |
| 144 |
During scalability testing, it is observed that synchronized actions |
| 145 |
among CPUs become challenging as the number of cores grows. This is |
| 146 |
also true for the ability of a system to enter package level C-states. |
| 147 |
|
| 148 |
To make sure the intel_powerclamp driver scales well, online |
| 149 |
calibration is implemented. The goals for doing such a calibration |
| 150 |
are: |
| 151 |
|
| 152 |
a) determine the effective range of idle injection ratio |
| 153 |
b) determine the amount of compensation needed at each target ratio |
| 154 |
|
| 155 |
Compensation to each target ratio consists of two parts: |
| 156 |
|
| 157 |
a) steady state error compensation |
| 158 |
|
| 159 |
This is to offset the error occurring when the system can |
| 160 |
enter idle without extra wakeups (such as external interrupts). |
| 161 |
|
| 162 |
b) dynamic error compensation |
| 163 |
|
| 164 |
When an excessive amount of wakeups occurs during idle, an |
| 165 |
additional idle ratio can be added to quiet interrupts, by |
| 166 |
slowing down CPU activities. |
| 167 |
|
| 168 |
A debugfs file is provided for the user to examine compensation |
| 169 |
progress and results, such as on a Westmere system:: |
| 170 |
|
| 171 |
[jacob@nex01 ~]$ cat |
| 172 |
/sys/kernel/debug/intel_powerclamp/powerclamp_calib |
| 173 |
controlling cpu: 0 |
| 174 |
pct confidence steady dynamic (compensation) |
| 175 |
0 0 0 0 |
| 176 |
1 1 0 0 |
| 177 |
2 1 1 0 |
| 178 |
3 3 1 0 |
| 179 |
4 3 1 0 |
| 180 |
5 3 1 0 |
| 181 |
6 3 1 0 |
| 182 |
7 3 1 0 |
| 183 |
8 3 1 0 |
| 184 |
... |
| 185 |
30 3 2 0 |
| 186 |
31 3 2 0 |
| 187 |
32 3 1 0 |
| 188 |
33 3 2 0 |
| 189 |
34 3 1 0 |
| 190 |
35 3 2 0 |
| 191 |
36 3 1 0 |
| 192 |
37 3 2 0 |
| 193 |
38 3 1 0 |
| 194 |
39 3 2 0 |
| 195 |
40 3 3 0 |
| 196 |
41 3 1 0 |
| 197 |
42 3 2 0 |
| 198 |
43 3 1 0 |
| 199 |
44 3 1 0 |
| 200 |
45 3 2 0 |
| 201 |
46 3 3 0 |
| 202 |
47 3 0 0 |
| 203 |
48 3 2 0 |
| 204 |
49 3 3 0 |
| 205 |
|
| 206 |
Calibration occurs during runtime. No offline method is available. |
| 207 |
Steady state compensation is used only when confidence levels of all |
| 208 |
adjacent ratios have reached satisfactory level. A confidence level |
| 209 |
is accumulated based on clean data collected at runtime. Data |
| 210 |
collected during a period without extra interrupts is considered |
| 211 |
clean. |
| 212 |
|
| 213 |
To compensate for excessive amounts of wakeup during idle, additional |
| 214 |
idle time is injected when such a condition is detected. Currently, |
| 215 |
we have a simple algorithm to double the injection ratio. A possible |
| 216 |
enhancement might be to throttle the offending IRQ, such as delaying |
| 217 |
EOI for level triggered interrupts. But it is a challenge to be |
| 218 |
non-intrusive to the scheduler or the IRQ core code. |
| 219 |
|
| 220 |
|
| 221 |
CPU Online/Offline |
| 222 |
------------------ |
| 223 |
Per-CPU kernel threads are started/stopped upon receiving |
| 224 |
notifications of CPU hotplug activities. The intel_powerclamp driver |
| 225 |
keeps track of clamping kernel threads, even after they are migrated |
| 226 |
to other CPUs, after a CPU offline event. |
| 227 |
|
| 228 |
|
| 229 |
Performance Analysis |
| 230 |
==================== |
| 231 |
This section describes the general performance data collected on |
| 232 |
multiple systems, including Westmere (80P) and Ivy Bridge (4P, 8P). |
| 233 |
|
| 234 |
Effectiveness and Limitations |
| 235 |
----------------------------- |
| 236 |
The maximum range that idle injection is allowed is capped at 50 |
| 237 |
percent. As mentioned earlier, since interrupts are allowed during |
| 238 |
forced idle time, excessive interrupts could result in less |
| 239 |
effectiveness. The extreme case would be doing a ping -f to generated |
| 240 |
flooded network interrupts without much CPU acknowledgement. In this |
| 241 |
case, little can be done from the idle injection threads. In most |
| 242 |
normal cases, such as scp a large file, applications can be throttled |
| 243 |
by the powerclamp driver, since slowing down the CPU also slows down |
| 244 |
network protocol processing, which in turn reduces interrupts. |
| 245 |
|
| 246 |
When control parameters change at runtime by the controlling CPU, it |
| 247 |
may take an additional period for the rest of the CPUs to catch up |
| 248 |
with the changes. During this time, idle injection is out of sync, |
| 249 |
thus not able to enter package C- states at the expected ratio. But |
| 250 |
this effect is minor, in that in most cases change to the target |
| 251 |
ratio is updated much less frequently than the idle injection |
| 252 |
frequency. |
| 253 |
|
| 254 |
Scalability |
| 255 |
----------- |
| 256 |
Tests also show a minor, but measurable, difference between the 4P/8P |
| 257 |
Ivy Bridge system and the 80P Westmere server under 50% idle ratio. |
| 258 |
More compensation is needed on Westmere for the same amount of |
| 259 |
target idle ratio. The compensation also increases as the idle ratio |
| 260 |
gets larger. The above reason constitutes the need for the |
| 261 |
calibration code. |
| 262 |
|
| 263 |
On the IVB 8P system, compared to an offline CPU, powerclamp can |
| 264 |
achieve up to 40% better performance per watt. (measured by a spin |
| 265 |
counter summed over per CPU counting threads spawned for all running |
| 266 |
CPUs). |
| 267 |
|
| 268 |
Usage and Interfaces |
| 269 |
==================== |
| 270 |
The powerclamp driver is registered to the generic thermal layer as a |
| 271 |
cooling device. Currently, it’s not bound to any thermal zones:: |
| 272 |
|
| 273 |
jacob@chromoly:/sys/class/thermal/cooling_device14$ grep . * |
| 274 |
cur_state:0 |
| 275 |
max_state:50 |
| 276 |
type:intel_powerclamp |
| 277 |
|
| 278 |
cur_state allows user to set the desired idle percentage. Writing 0 to |
| 279 |
cur_state will stop idle injection. Writing a value between 1 and |
| 280 |
max_state will start the idle injection. Reading cur_state returns the |
| 281 |
actual and current idle percentage. This may not be the same value |
| 282 |
set by the user in that current idle percentage depends on workload |
| 283 |
and includes natural idle. When idle injection is disabled, reading |
| 284 |
cur_state returns value -1 instead of 0 which is to avoid confusing |
| 285 |
100% busy state with the disabled state. |
| 286 |
|
| 287 |
Example usage: |
| 288 |
|
| 289 |
- To inject 25% idle time:: |
| 290 |
|
| 291 |
$ sudo sh -c "echo 25 > /sys/class/thermal/cooling_device80/cur_state |
| 292 |
|
| 293 |
If the system is not busy and has more than 25% idle time already, |
| 294 |
then the powerclamp driver will not start idle injection. Using Top |
| 295 |
will not show idle injection kernel threads. |
| 296 |
|
| 297 |
If the system is busy (spin test below) and has less than 25% natural |
| 298 |
idle time, powerclamp kernel threads will do idle injection. Forced |
| 299 |
idle time is accounted as normal idle in that common code path is |
| 300 |
taken as the idle task. |
| 301 |
|
| 302 |
In this example, 24.1% idle is shown. This helps the system admin or |
| 303 |
user determine the cause of slowdown, when a powerclamp driver is in action:: |
| 304 |
|
| 305 |
|
| 306 |
Tasks: 197 total, 1 running, 196 sleeping, 0 stopped, 0 zombie |
| 307 |
Cpu(s): 71.2%us, 4.7%sy, 0.0%ni, 24.1%id, 0.0%wa, 0.0%hi, 0.0%si, 0.0%st |
| 308 |
Mem: 3943228k total, 1689632k used, 2253596k free, 74960k buffers |
| 309 |
Swap: 4087804k total, 0k used, 4087804k free, 945336k cached |
| 310 |
|
| 311 |
PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND |
| 312 |
3352 jacob 20 0 262m 644 428 S 286 0.0 0:17.16 spin |
| 313 |
3341 root -51 0 0 0 0 D 25 0.0 0:01.62 kidle_inject/0 |
| 314 |
3344 root -51 0 0 0 0 D 25 0.0 0:01.60 kidle_inject/3 |
| 315 |
3342 root -51 0 0 0 0 D 25 0.0 0:01.61 kidle_inject/1 |
| 316 |
3343 root -51 0 0 0 0 D 25 0.0 0:01.60 kidle_inject/2 |
| 317 |
2935 jacob 20 0 696m 125m 35m S 5 3.3 0:31.11 firefox |
| 318 |
1546 root 20 0 158m 20m 6640 S 3 0.5 0:26.97 Xorg |
| 319 |
2100 jacob 20 0 1223m 88m 30m S 3 2.3 0:23.68 compiz |
| 320 |
|
| 321 |
Tests have shown that by using the powerclamp driver as a cooling |
| 322 |
device, a PID based userspace thermal controller can manage to |
| 323 |
control CPU temperature effectively, when no other thermal influence |
| 324 |
is added. For example, a UltraBook user can compile the kernel under |
| 325 |
certain temperature (below most active trip points). |
| 326 |
|
| 327 |
Module Parameters |
| 328 |
================= |
| 329 |
|
| 330 |
``cpumask`` (RW) |
| 331 |
A bit mask of CPUs to inject idle. The format of the bitmask is same as |
| 332 |
used in other subsystems like in /proc/irq/\*/smp_affinity. The mask is |
| 333 |
comma separated 32 bit groups. Each CPU is one bit. For example for a 256 |
| 334 |
CPU system the full mask is: |
| 335 |
ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff |
| 336 |
|
| 337 |
The rightmost mask is for CPU 0-32. |
| 338 |
|
| 339 |
``max_idle`` (RW) |
| 340 |
Maximum injected idle time to the total CPU time ratio in percent range |
| 341 |
from 1 to 100. Even if the cooling device max_state is always 100 (100%), |
| 342 |
this parameter allows to add a max idle percent limit. The default is 50, |
| 343 |
to match the current implementation of powerclamp driver. Also doesn't |
| 344 |
allow value more than 75, if the cpumask includes every CPU present in |
| 345 |
the system. |