1 =======================
2 INTEL POWERCLAMP DRIVER
3 =======================
4 By: Arjan van de Ven <arjan@linux.intel.com>
5 Jacob Pan <jacob.jun.pan@linux.intel.com>
11 (*) Theory of Operation
15 (*) Performance Analysis
16 - Effectiveness and Limitations
17 - Power vs Performance
20 - Comparison with Alternative Techniques
22 (*) Usage and Interfaces
23 - Generic Thermal Layer (sysfs)
30 Consider the situation where a system’s power consumption must be
31 reduced at runtime, due to power budget, thermal constraint, or noise
32 level, and where active cooling is not preferred. Software managed
33 passive power reduction must be performed to prevent the hardware
34 actions that are designed for catastrophic scenarios.
36 Currently, P-states, T-states (clock modulation), and CPU offlining
37 are used for CPU throttling.
39 On Intel CPUs, C-states provide effective power reduction, but so far
40 they’re only used opportunistically, based on workload. With the
41 development of intel_powerclamp driver, the method of synchronizing
42 idle injection across all online CPU threads was introduced. The goal
43 is to achieve forced and controllable C-state residency.
45 Test/Analysis has been made in the areas of power, performance,
46 scalability, and user experience. In many cases, clear advantage is
47 shown over taking the CPU offline or modulating the CPU clock.
57 On modern Intel processors (Nehalem or later), package level C-state
58 residency is available in MSRs, thus also available to the kernel.
61 #define MSR_PKG_C2_RESIDENCY 0x60D
62 #define MSR_PKG_C3_RESIDENCY 0x3F8
63 #define MSR_PKG_C6_RESIDENCY 0x3F9
64 #define MSR_PKG_C7_RESIDENCY 0x3FA
66 If the kernel can also inject idle time to the system, then a
67 closed-loop control system can be established that manages package
68 level C-state. The intel_powerclamp driver is conceived as such a
69 control system, where the target set point is a user-selected idle
70 ratio (based on power reduction), and the error is the difference
71 between the actual package level C-state residency ratio and the target idle
74 Injection is controlled by high priority kernel threads, spawned for
77 These kernel threads, with SCHED_FIFO class, are created to perform
78 clamping actions of controlled duty ratio and duration. Each per-CPU
79 thread synchronizes its idle time and duration, based on the rounding
80 of jiffies, so accumulated errors can be prevented to avoid a jittery
81 effect. Threads are also bound to the CPU such that they cannot be
82 migrated, unless the CPU is taken offline. In this case, threads
83 belong to the offlined CPUs will be terminated immediately.
85 Running as SCHED_FIFO and relatively high priority, also allows such
86 scheme to work for both preemptable and non-preemptable kernels.
87 Alignment of idle time around jiffies ensures scalability for HZ
88 values. This effect can be better visualized using a Perf timechart.
89 The following diagram shows the behavior of kernel thread
90 kidle_inject/cpu. During idle injection, it runs monitor/mwait idle
91 for a given "duration", then relinquishes the CPU to other tasks,
92 until the next time interval.
94 The NOHZ schedule tick is disabled during idle time, but interrupts
95 are not masked. Tests show that the extra wakeups from scheduler tick
96 have a dramatic impact on the effectiveness of the powerclamp driver
97 on large scale systems (Westmere system with 80 processors).
100 ____________ ____________
101 kidle_inject/0 | sleep | mwait | sleep |
102 _________| |________| |_______
105 ____________ ____________
106 kidle_inject/1 | sleep | mwait | sleep |
107 _________| |________| |_______
111 roundup(jiffies, interval)
113 Only one CPU is allowed to collect statistics and update global
114 control parameters. This CPU is referred to as the controlling CPU in
115 this document. The controlling CPU is elected at runtime, with a
116 policy that favors BSP, taking into account the possibility of a CPU
119 In terms of dynamics of the idle control system, package level idle
120 time is considered largely as a non-causal system where its behavior
121 cannot be based on the past or current input. Therefore, the
122 intel_powerclamp driver attempts to enforce the desired idle time
123 instantly as given input (target idle ratio). After injection,
124 powerclamp monitors the actual idle for a given time window and adjust
125 the next injection accordingly to avoid over/under correction.
127 When used in a causal control system, such as a temperature control,
128 it is up to the user of this driver to implement algorithms where
129 past samples and outputs are included in the feedback. For example, a
130 PID-based thermal controller can use the powerclamp driver to
131 maintain a desired target temperature, based on integral and
132 derivative gains of the past samples.
138 During scalability testing, it is observed that synchronized actions
139 among CPUs become challenging as the number of cores grows. This is
140 also true for the ability of a system to enter package level C-states.
142 To make sure the intel_powerclamp driver scales well, online
143 calibration is implemented. The goals for doing such a calibration
146 a) determine the effective range of idle injection ratio
147 b) determine the amount of compensation needed at each target ratio
149 Compensation to each target ratio consists of two parts:
151 a) steady state error compensation
152 This is to offset the error occurring when the system can
153 enter idle without extra wakeups (such as external interrupts).
155 b) dynamic error compensation
156 When an excessive amount of wakeups occurs during idle, an
157 additional idle ratio can be added to quiet interrupts, by
158 slowing down CPU activities.
160 A debugfs file is provided for the user to examine compensation
161 progress and results, such as on a Westmere system.
163 /sys/kernel/debug/intel_powerclamp/powerclamp_calib
165 pct confidence steady dynamic (compensation)
197 Calibration occurs during runtime. No offline method is available.
198 Steady state compensation is used only when confidence levels of all
199 adjacent ratios have reached satisfactory level. A confidence level
200 is accumulated based on clean data collected at runtime. Data
201 collected during a period without extra interrupts is considered
204 To compensate for excessive amounts of wakeup during idle, additional
205 idle time is injected when such a condition is detected. Currently,
206 we have a simple algorithm to double the injection ratio. A possible
207 enhancement might be to throttle the offending IRQ, such as delaying
208 EOI for level triggered interrupts. But it is a challenge to be
209 non-intrusive to the scheduler or the IRQ core code.
214 Per-CPU kernel threads are started/stopped upon receiving
215 notifications of CPU hotplug activities. The intel_powerclamp driver
216 keeps track of clamping kernel threads, even after they are migrated
217 to other CPUs, after a CPU offline event.
220 =====================
222 =====================
223 This section describes the general performance data collected on
224 multiple systems, including Westmere (80P) and Ivy Bridge (4P, 8P).
226 Effectiveness and Limitations
227 -----------------------------
228 The maximum range that idle injection is allowed is capped at 50
229 percent. As mentioned earlier, since interrupts are allowed during
230 forced idle time, excessive interrupts could result in less
231 effectiveness. The extreme case would be doing a ping -f to generated
232 flooded network interrupts without much CPU acknowledgement. In this
233 case, little can be done from the idle injection threads. In most
234 normal cases, such as scp a large file, applications can be throttled
235 by the powerclamp driver, since slowing down the CPU also slows down
236 network protocol processing, which in turn reduces interrupts.
238 When control parameters change at runtime by the controlling CPU, it
239 may take an additional period for the rest of the CPUs to catch up
240 with the changes. During this time, idle injection is out of sync,
241 thus not able to enter package C- states at the expected ratio. But
242 this effect is minor, in that in most cases change to the target
243 ratio is updated much less frequently than the idle injection
248 Tests also show a minor, but measurable, difference between the 4P/8P
249 Ivy Bridge system and the 80P Westmere server under 50% idle ratio.
250 More compensation is needed on Westmere for the same amount of
251 target idle ratio. The compensation also increases as the idle ratio
252 gets larger. The above reason constitutes the need for the
255 On the IVB 8P system, compared to an offline CPU, powerclamp can
256 achieve up to 40% better performance per watt. (measured by a spin
257 counter summed over per CPU counting threads spawned for all running
263 The powerclamp driver is registered to the generic thermal layer as a
264 cooling device. Currently, it’s not bound to any thermal zones.
266 jacob@chromoly:/sys/class/thermal/cooling_device14$ grep . *
269 type:intel_powerclamp
271 cur_state allows user to set the desired idle percentage. Writing 0 to
272 cur_state will stop idle injection. Writing a value between 1 and
273 max_state will start the idle injection. Reading cur_state returns the
274 actual and current idle percentage. This may not be the same value
275 set by the user in that current idle percentage depends on workload
276 and includes natural idle. When idle injection is disabled, reading
277 cur_state returns value -1 instead of 0 which is to avoid confusing
278 100% busy state with the disabled state.
281 - To inject 25% idle time
282 $ sudo sh -c "echo 25 > /sys/class/thermal/cooling_device80/cur_state
285 If the system is not busy and has more than 25% idle time already,
286 then the powerclamp driver will not start idle injection. Using Top
287 will not show idle injection kernel threads.
289 If the system is busy (spin test below) and has less than 25% natural
290 idle time, powerclamp kernel threads will do idle injection. Forced
291 idle time is accounted as normal idle in that common code path is
292 taken as the idle task.
294 In this example, 24.1% idle is shown. This helps the system admin or
295 user determine the cause of slowdown, when a powerclamp driver is in action.
298 Tasks: 197 total, 1 running, 196 sleeping, 0 stopped, 0 zombie
299 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
300 Mem: 3943228k total, 1689632k used, 2253596k free, 74960k buffers
301 Swap: 4087804k total, 0k used, 4087804k free, 945336k cached
303 PID USER PR NI VIRT RES SHR S %CPU %MEM TIME+ COMMAND
304 3352 jacob 20 0 262m 644 428 S 286 0.0 0:17.16 spin
305 3341 root -51 0 0 0 0 D 25 0.0 0:01.62 kidle_inject/0
306 3344 root -51 0 0 0 0 D 25 0.0 0:01.60 kidle_inject/3
307 3342 root -51 0 0 0 0 D 25 0.0 0:01.61 kidle_inject/1
308 3343 root -51 0 0 0 0 D 25 0.0 0:01.60 kidle_inject/2
309 2935 jacob 20 0 696m 125m 35m S 5 3.3 0:31.11 firefox
310 1546 root 20 0 158m 20m 6640 S 3 0.5 0:26.97 Xorg
311 2100 jacob 20 0 1223m 88m 30m S 3 2.3 0:23.68 compiz
313 Tests have shown that by using the powerclamp driver as a cooling
314 device, a PID based userspace thermal controller can manage to
315 control CPU temperature effectively, when no other thermal influence
316 is added. For example, a UltraBook user can compile the kernel under
317 certain temperature (below most active trip points).