aboutsummaryrefslogtreecommitdiffstats
path: root/drivers/cpuidle
diff options
context:
space:
mode:
authorArjan van de Ven <arjan@infradead.org>2009-09-21 20:04:08 -0400
committerLinus Torvalds <torvalds@linux-foundation.org>2009-09-22 10:17:45 -0400
commit69d25870f20c4b2563304f2b79c5300dd60a067e (patch)
treecda2b2d65c1be95420c6ba92ae2d40fade4232c4 /drivers/cpuidle
parent45d80eea87c9f8292d2d33173d6866c0ec57238a (diff)
cpuidle: fix the menu governor to boost IO performance
Fix the menu idle governor which balances power savings, energy efficiency and performance impact. The reason for a reworked governor is that there have been serious performance issues reported with the existing code on Nehalem server systems. To show this I'm sure Andrew wants to see benchmark results: (benchmark is "fio", "no cstates" is using "idle=poll") no cstates current linux new algorithm 1 disk 107 Mb/s 85 Mb/s 105 Mb/s 2 disks 215 Mb/s 123 Mb/s 209 Mb/s 12 disks 590 Mb/s 320 Mb/s 585 Mb/s In various power benchmark measurements, no degredation was found by our measurement&diagnostics team. Obviously a small percentage more power was used in the "fio" benchmark, due to the much higher performance. While it would be a novel idea to describe the new algorithm in this commit message, I cheaped out and described it in comments in the code instead. [changes since first post: spelling fixes from akpm, review feedback, folded menu-tng into menu.c] Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Venkatesh Pallipadi <venkatesh.pallipadi@intel.com> Cc: Len Brown <lenb@kernel.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Yanmin Zhang <yanmin_zhang@linux.intel.com> Acked-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Diffstat (limited to 'drivers/cpuidle')
-rw-r--r--drivers/cpuidle/governors/menu.c251
1 files changed, 212 insertions, 39 deletions
diff --git a/drivers/cpuidle/governors/menu.c b/drivers/cpuidle/governors/menu.c
index f1df59f59a37..9f3d77532ab9 100644
--- a/drivers/cpuidle/governors/menu.c
+++ b/drivers/cpuidle/governors/menu.c
@@ -2,8 +2,12 @@
2 * menu.c - the menu idle governor 2 * menu.c - the menu idle governor
3 * 3 *
4 * Copyright (C) 2006-2007 Adam Belay <abelay@novell.com> 4 * Copyright (C) 2006-2007 Adam Belay <abelay@novell.com>
5 * Copyright (C) 2009 Intel Corporation
6 * Author:
7 * Arjan van de Ven <arjan@linux.intel.com>
5 * 8 *
6 * This code is licenced under the GPL. 9 * This code is licenced under the GPL version 2 as described
10 * in the COPYING file that acompanies the Linux Kernel.
7 */ 11 */
8 12
9#include <linux/kernel.h> 13#include <linux/kernel.h>
@@ -13,20 +17,153 @@
13#include <linux/ktime.h> 17#include <linux/ktime.h>
14#include <linux/hrtimer.h> 18#include <linux/hrtimer.h>
15#include <linux/tick.h> 19#include <linux/tick.h>
20#include <linux/sched.h>
16 21
17#define BREAK_FUZZ 4 /* 4 us */ 22#define BUCKETS 12
18#define PRED_HISTORY_PCT 50 23#define RESOLUTION 1024
24#define DECAY 4
25#define MAX_INTERESTING 50000
26
27/*
28 * Concepts and ideas behind the menu governor
29 *
30 * For the menu governor, there are 3 decision factors for picking a C
31 * state:
32 * 1) Energy break even point
33 * 2) Performance impact
34 * 3) Latency tolerance (from pmqos infrastructure)
35 * These these three factors are treated independently.
36 *
37 * Energy break even point
38 * -----------------------
39 * C state entry and exit have an energy cost, and a certain amount of time in
40 * the C state is required to actually break even on this cost. CPUIDLE
41 * provides us this duration in the "target_residency" field. So all that we
42 * need is a good prediction of how long we'll be idle. Like the traditional
43 * menu governor, we start with the actual known "next timer event" time.
44 *
45 * Since there are other source of wakeups (interrupts for example) than
46 * the next timer event, this estimation is rather optimistic. To get a
47 * more realistic estimate, a correction factor is applied to the estimate,
48 * that is based on historic behavior. For example, if in the past the actual
49 * duration always was 50% of the next timer tick, the correction factor will
50 * be 0.5.
51 *
52 * menu uses a running average for this correction factor, however it uses a
53 * set of factors, not just a single factor. This stems from the realization
54 * that the ratio is dependent on the order of magnitude of the expected
55 * duration; if we expect 500 milliseconds of idle time the likelihood of
56 * getting an interrupt very early is much higher than if we expect 50 micro
57 * seconds of idle time. A second independent factor that has big impact on
58 * the actual factor is if there is (disk) IO outstanding or not.
59 * (as a special twist, we consider every sleep longer than 50 milliseconds
60 * as perfect; there are no power gains for sleeping longer than this)
61 *
62 * For these two reasons we keep an array of 12 independent factors, that gets
63 * indexed based on the magnitude of the expected duration as well as the
64 * "is IO outstanding" property.
65 *
66 * Limiting Performance Impact
67 * ---------------------------
68 * C states, especially those with large exit latencies, can have a real
69 * noticable impact on workloads, which is not acceptable for most sysadmins,
70 * and in addition, less performance has a power price of its own.
71 *
72 * As a general rule of thumb, menu assumes that the following heuristic
73 * holds:
74 * The busier the system, the less impact of C states is acceptable
75 *
76 * This rule-of-thumb is implemented using a performance-multiplier:
77 * If the exit latency times the performance multiplier is longer than
78 * the predicted duration, the C state is not considered a candidate
79 * for selection due to a too high performance impact. So the higher
80 * this multiplier is, the longer we need to be idle to pick a deep C
81 * state, and thus the less likely a busy CPU will hit such a deep
82 * C state.
83 *
84 * Two factors are used in determing this multiplier:
85 * a value of 10 is added for each point of "per cpu load average" we have.
86 * a value of 5 points is added for each process that is waiting for
87 * IO on this CPU.
88 * (these values are experimentally determined)
89 *
90 * The load average factor gives a longer term (few seconds) input to the
91 * decision, while the iowait value gives a cpu local instantanious input.
92 * The iowait factor may look low, but realize that this is also already
93 * represented in the system load average.
94 *
95 */
19 96
20struct menu_device { 97struct menu_device {
21 int last_state_idx; 98 int last_state_idx;
22 99
23 unsigned int expected_us; 100 unsigned int expected_us;
24 unsigned int predicted_us; 101 u64 predicted_us;
25 unsigned int current_predicted_us; 102 unsigned int measured_us;
26 unsigned int last_measured_us; 103 unsigned int exit_us;
27 unsigned int elapsed_us; 104 unsigned int bucket;
105 u64 correction_factor[BUCKETS];
28}; 106};
29 107
108
109#define LOAD_INT(x) ((x) >> FSHIFT)
110#define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
111
112static int get_loadavg(void)
113{
114 unsigned long this = this_cpu_load();
115
116
117 return LOAD_INT(this) * 10 + LOAD_FRAC(this) / 10;
118}
119
120static inline int which_bucket(unsigned int duration)
121{
122 int bucket = 0;
123
124 /*
125 * We keep two groups of stats; one with no
126 * IO pending, one without.
127 * This allows us to calculate
128 * E(duration)|iowait
129 */
130 if (nr_iowait_cpu())
131 bucket = BUCKETS/2;
132
133 if (duration < 10)
134 return bucket;
135 if (duration < 100)
136 return bucket + 1;
137 if (duration < 1000)
138 return bucket + 2;
139 if (duration < 10000)
140 return bucket + 3;
141 if (duration < 100000)
142 return bucket + 4;
143 return bucket + 5;
144}
145
146/*
147 * Return a multiplier for the exit latency that is intended
148 * to take performance requirements into account.
149 * The more performance critical we estimate the system
150 * to be, the higher this multiplier, and thus the higher
151 * the barrier to go to an expensive C state.
152 */
153static inline int performance_multiplier(void)
154{
155 int mult = 1;
156
157 /* for higher loadavg, we are more reluctant */
158
159 mult += 2 * get_loadavg();
160
161 /* for IO wait tasks (per cpu!) we add 5x each */
162 mult += 10 * nr_iowait_cpu();
163
164 return mult;
165}
166
30static DEFINE_PER_CPU(struct menu_device, menu_devices); 167static DEFINE_PER_CPU(struct menu_device, menu_devices);
31 168
32/** 169/**
@@ -38,37 +175,59 @@ static int menu_select(struct cpuidle_device *dev)
38 struct menu_device *data = &__get_cpu_var(menu_devices); 175 struct menu_device *data = &__get_cpu_var(menu_devices);
39 int latency_req = pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY); 176 int latency_req = pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY);
40 int i; 177 int i;
178 int multiplier;
179
180 data->last_state_idx = 0;
181 data->exit_us = 0;
41 182
42 /* Special case when user has set very strict latency requirement */ 183 /* Special case when user has set very strict latency requirement */
43 if (unlikely(latency_req == 0)) { 184 if (unlikely(latency_req == 0))
44 data->last_state_idx = 0;
45 return 0; 185 return 0;
46 }
47 186
48 /* determine the expected residency time */ 187 /* determine the expected residency time, round up */
49 data->expected_us = 188 data->expected_us =
50 (u32) ktime_to_ns(tick_nohz_get_sleep_length()) / 1000; 189 DIV_ROUND_UP((u32)ktime_to_ns(tick_nohz_get_sleep_length()), 1000);
190
191
192 data->bucket = which_bucket(data->expected_us);
193
194 multiplier = performance_multiplier();
195
196 /*
197 * if the correction factor is 0 (eg first time init or cpu hotplug
198 * etc), we actually want to start out with a unity factor.
199 */
200 if (data->correction_factor[data->bucket] == 0)
201 data->correction_factor[data->bucket] = RESOLUTION * DECAY;
202
203 /* Make sure to round up for half microseconds */
204 data->predicted_us = DIV_ROUND_CLOSEST(
205 data->expected_us * data->correction_factor[data->bucket],
206 RESOLUTION * DECAY);
207
208 /*
209 * We want to default to C1 (hlt), not to busy polling
210 * unless the timer is happening really really soon.
211 */
212 if (data->expected_us > 5)
213 data->last_state_idx = CPUIDLE_DRIVER_STATE_START;
51 214
52 /* Recalculate predicted_us based on prediction_history_pct */
53 data->predicted_us *= PRED_HISTORY_PCT;
54 data->predicted_us += (100 - PRED_HISTORY_PCT) *
55 data->current_predicted_us;
56 data->predicted_us /= 100;
57 215
58 /* find the deepest idle state that satisfies our constraints */ 216 /* find the deepest idle state that satisfies our constraints */
59 for (i = CPUIDLE_DRIVER_STATE_START + 1; i < dev->state_count; i++) { 217 for (i = CPUIDLE_DRIVER_STATE_START; i < dev->state_count; i++) {
60 struct cpuidle_state *s = &dev->states[i]; 218 struct cpuidle_state *s = &dev->states[i];
61 219
62 if (s->target_residency > data->expected_us)
63 break;
64 if (s->target_residency > data->predicted_us) 220 if (s->target_residency > data->predicted_us)
65 break; 221 break;
66 if (s->exit_latency > latency_req) 222 if (s->exit_latency > latency_req)
67 break; 223 break;
224 if (s->exit_latency * multiplier > data->predicted_us)
225 break;
226 data->exit_us = s->exit_latency;
227 data->last_state_idx = i;
68 } 228 }
69 229
70 data->last_state_idx = i - 1; 230 return data->last_state_idx;
71 return i - 1;
72} 231}
73 232
74/** 233/**
@@ -85,35 +244,49 @@ static void menu_reflect(struct cpuidle_device *dev)
85 unsigned int last_idle_us = cpuidle_get_last_residency(dev); 244 unsigned int last_idle_us = cpuidle_get_last_residency(dev);
86 struct cpuidle_state *target = &dev->states[last_idx]; 245 struct cpuidle_state *target = &dev->states[last_idx];
87 unsigned int measured_us; 246 unsigned int measured_us;
247 u64 new_factor;
88 248
89 /* 249 /*
90 * Ugh, this idle state doesn't support residency measurements, so we 250 * Ugh, this idle state doesn't support residency measurements, so we
91 * are basically lost in the dark. As a compromise, assume we slept 251 * are basically lost in the dark. As a compromise, assume we slept
92 * for one full standard timer tick. However, be aware that this 252 * for the whole expected time.
93 * could potentially result in a suboptimal state transition.
94 */ 253 */
95 if (unlikely(!(target->flags & CPUIDLE_FLAG_TIME_VALID))) 254 if (unlikely(!(target->flags & CPUIDLE_FLAG_TIME_VALID)))
96 last_idle_us = USEC_PER_SEC / HZ; 255 last_idle_us = data->expected_us;
256
257
258 measured_us = last_idle_us;
97 259
98 /* 260 /*
99 * measured_us and elapsed_us are the cumulative idle time, since the 261 * We correct for the exit latency; we are assuming here that the
100 * last time we were woken out of idle by an interrupt. 262 * exit latency happens after the event that we're interested in.
101 */ 263 */
102 if (data->elapsed_us <= data->elapsed_us + last_idle_us) 264 if (measured_us > data->exit_us)
103 measured_us = data->elapsed_us + last_idle_us; 265 measured_us -= data->exit_us;
266
267
268 /* update our correction ratio */
269
270 new_factor = data->correction_factor[data->bucket]
271 * (DECAY - 1) / DECAY;
272
273 if (data->expected_us > 0 && data->measured_us < MAX_INTERESTING)
274 new_factor += RESOLUTION * measured_us / data->expected_us;
104 else 275 else
105 measured_us = -1; 276 /*
277 * we were idle so long that we count it as a perfect
278 * prediction
279 */
280 new_factor += RESOLUTION;
106 281
107 /* Predict time until next break event */ 282 /*
108 data->current_predicted_us = max(measured_us, data->last_measured_us); 283 * We don't want 0 as factor; we always want at least
284 * a tiny bit of estimated time.
285 */
286 if (new_factor == 0)
287 new_factor = 1;
109 288
110 if (last_idle_us + BREAK_FUZZ < 289 data->correction_factor[data->bucket] = new_factor;
111 data->expected_us - target->exit_latency) {
112 data->last_measured_us = measured_us;
113 data->elapsed_us = 0;
114 } else {
115 data->elapsed_us = measured_us;
116 }
117} 290}
118 291
119/** 292/**