diff options
Diffstat (limited to 'Documentation/scheduler/sched-stats.txt')
-rw-r--r-- | Documentation/scheduler/sched-stats.txt | 156 |
1 files changed, 156 insertions, 0 deletions
diff --git a/Documentation/scheduler/sched-stats.txt b/Documentation/scheduler/sched-stats.txt new file mode 100644 index 000000000000..442e14d35dea --- /dev/null +++ b/Documentation/scheduler/sched-stats.txt | |||
@@ -0,0 +1,156 @@ | |||
1 | Version 14 of schedstats includes support for sched_domains, which hit the | ||
2 | mainline kernel in 2.6.20 although it is identical to the stats from version | ||
3 | 12 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel | ||
4 | release). Some counters make more sense to be per-runqueue; other to be | ||
5 | per-domain. Note that domains (and their associated information) will only | ||
6 | be pertinent and available on machines utilizing CONFIG_SMP. | ||
7 | |||
8 | In version 14 of schedstat, there is at least one level of domain | ||
9 | statistics for each cpu listed, and there may well be more than one | ||
10 | domain. Domains have no particular names in this implementation, but | ||
11 | the highest numbered one typically arbitrates balancing across all the | ||
12 | cpus on the machine, while domain0 is the most tightly focused domain, | ||
13 | sometimes balancing only between pairs of cpus. At this time, there | ||
14 | are no architectures which need more than three domain levels. The first | ||
15 | field in the domain stats is a bit map indicating which cpus are affected | ||
16 | by that domain. | ||
17 | |||
18 | These fields are counters, and only increment. Programs which make use | ||
19 | of these will need to start with a baseline observation and then calculate | ||
20 | the change in the counters at each subsequent observation. A perl script | ||
21 | which does this for many of the fields is available at | ||
22 | |||
23 | http://eaglet.rain.com/rick/linux/schedstat/ | ||
24 | |||
25 | Note that any such script will necessarily be version-specific, as the main | ||
26 | reason to change versions is changes in the output format. For those wishing | ||
27 | to write their own scripts, the fields are described here. | ||
28 | |||
29 | CPU statistics | ||
30 | -------------- | ||
31 | cpu<N> 1 2 3 4 5 6 7 8 9 10 11 12 | ||
32 | |||
33 | NOTE: In the sched_yield() statistics, the active queue is considered empty | ||
34 | if it has only one process in it, since obviously the process calling | ||
35 | sched_yield() is that process. | ||
36 | |||
37 | First four fields are sched_yield() statistics: | ||
38 | 1) # of times both the active and the expired queue were empty | ||
39 | 2) # of times just the active queue was empty | ||
40 | 3) # of times just the expired queue was empty | ||
41 | 4) # of times sched_yield() was called | ||
42 | |||
43 | Next three are schedule() statistics: | ||
44 | 5) # of times we switched to the expired queue and reused it | ||
45 | 6) # of times schedule() was called | ||
46 | 7) # of times schedule() left the processor idle | ||
47 | |||
48 | Next two are try_to_wake_up() statistics: | ||
49 | 8) # of times try_to_wake_up() was called | ||
50 | 9) # of times try_to_wake_up() was called to wake up the local cpu | ||
51 | |||
52 | Next three are statistics describing scheduling latency: | ||
53 | 10) sum of all time spent running by tasks on this processor (in jiffies) | ||
54 | 11) sum of all time spent waiting to run by tasks on this processor (in | ||
55 | jiffies) | ||
56 | 12) # of timeslices run on this cpu | ||
57 | |||
58 | |||
59 | Domain statistics | ||
60 | ----------------- | ||
61 | One of these is produced per domain for each cpu described. (Note that if | ||
62 | CONFIG_SMP is not defined, *no* domains are utilized and these lines | ||
63 | will not appear in the output.) | ||
64 | |||
65 | domain<N> <cpumask> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 | ||
66 | |||
67 | The first field is a bit mask indicating what cpus this domain operates over. | ||
68 | |||
69 | The next 24 are a variety of load_balance() statistics in grouped into types | ||
70 | of idleness (idle, busy, and newly idle): | ||
71 | |||
72 | 1) # of times in this domain load_balance() was called when the | ||
73 | cpu was idle | ||
74 | 2) # of times in this domain load_balance() checked but found | ||
75 | the load did not require balancing when the cpu was idle | ||
76 | 3) # of times in this domain load_balance() tried to move one or | ||
77 | more tasks and failed, when the cpu was idle | ||
78 | 4) sum of imbalances discovered (if any) with each call to | ||
79 | load_balance() in this domain when the cpu was idle | ||
80 | 5) # of times in this domain pull_task() was called when the cpu | ||
81 | was idle | ||
82 | 6) # of times in this domain pull_task() was called even though | ||
83 | the target task was cache-hot when idle | ||
84 | 7) # of times in this domain load_balance() was called but did | ||
85 | not find a busier queue while the cpu was idle | ||
86 | 8) # of times in this domain a busier queue was found while the | ||
87 | cpu was idle but no busier group was found | ||
88 | |||
89 | 9) # of times in this domain load_balance() was called when the | ||
90 | cpu was busy | ||
91 | 10) # of times in this domain load_balance() checked but found the | ||
92 | load did not require balancing when busy | ||
93 | 11) # of times in this domain load_balance() tried to move one or | ||
94 | more tasks and failed, when the cpu was busy | ||
95 | 12) sum of imbalances discovered (if any) with each call to | ||
96 | load_balance() in this domain when the cpu was busy | ||
97 | 13) # of times in this domain pull_task() was called when busy | ||
98 | 14) # of times in this domain pull_task() was called even though the | ||
99 | target task was cache-hot when busy | ||
100 | 15) # of times in this domain load_balance() was called but did not | ||
101 | find a busier queue while the cpu was busy | ||
102 | 16) # of times in this domain a busier queue was found while the cpu | ||
103 | was busy but no busier group was found | ||
104 | |||
105 | 17) # of times in this domain load_balance() was called when the | ||
106 | cpu was just becoming idle | ||
107 | 18) # of times in this domain load_balance() checked but found the | ||
108 | load did not require balancing when the cpu was just becoming idle | ||
109 | 19) # of times in this domain load_balance() tried to move one or more | ||
110 | tasks and failed, when the cpu was just becoming idle | ||
111 | 20) sum of imbalances discovered (if any) with each call to | ||
112 | load_balance() in this domain when the cpu was just becoming idle | ||
113 | 21) # of times in this domain pull_task() was called when newly idle | ||
114 | 22) # of times in this domain pull_task() was called even though the | ||
115 | target task was cache-hot when just becoming idle | ||
116 | 23) # of times in this domain load_balance() was called but did not | ||
117 | find a busier queue while the cpu was just becoming idle | ||
118 | 24) # of times in this domain a busier queue was found while the cpu | ||
119 | was just becoming idle but no busier group was found | ||
120 | |||
121 | Next three are active_load_balance() statistics: | ||
122 | 25) # of times active_load_balance() was called | ||
123 | 26) # of times active_load_balance() tried to move a task and failed | ||
124 | 27) # of times active_load_balance() successfully moved a task | ||
125 | |||
126 | Next three are sched_balance_exec() statistics: | ||
127 | 28) sbe_cnt is not used | ||
128 | 29) sbe_balanced is not used | ||
129 | 30) sbe_pushed is not used | ||
130 | |||
131 | Next three are sched_balance_fork() statistics: | ||
132 | 31) sbf_cnt is not used | ||
133 | 32) sbf_balanced is not used | ||
134 | 33) sbf_pushed is not used | ||
135 | |||
136 | Next three are try_to_wake_up() statistics: | ||
137 | 34) # of times in this domain try_to_wake_up() awoke a task that | ||
138 | last ran on a different cpu in this domain | ||
139 | 35) # of times in this domain try_to_wake_up() moved a task to the | ||
140 | waking cpu because it was cache-cold on its own cpu anyway | ||
141 | 36) # of times in this domain try_to_wake_up() started passive balancing | ||
142 | |||
143 | /proc/<pid>/schedstat | ||
144 | ---------------- | ||
145 | schedstats also adds a new /proc/<pid/schedstat file to include some of | ||
146 | the same information on a per-process level. There are three fields in | ||
147 | this file correlating for that process to: | ||
148 | 1) time spent on the cpu | ||
149 | 2) time spent waiting on a runqueue | ||
150 | 3) # of timeslices run on this cpu | ||
151 | |||
152 | A program could be easily written to make use of these extra fields to | ||
153 | report on how well a particular process or set of processes is faring | ||
154 | under the scheduler's policies. A simple version of such a program is | ||
155 | available at | ||
156 | http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c | ||