Commit | Line | Data |
---|---|---|
b762f3ff JD |
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 | |
1da177e4 LT |
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 | -------------- | |
b762f3ff | 31 | cpu<N> 1 2 3 4 5 6 7 8 9 10 11 12 |
1da177e4 LT |
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 | ||
b762f3ff JD |
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 | |
1da177e4 | 47 | |
b762f3ff JD |
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 | |
1da177e4 LT |
51 | |
52 | Next three are statistics describing scheduling latency: | |
b762f3ff JD |
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 | |
1da177e4 LT |
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 | ||
b762f3ff | 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 |
1da177e4 LT |
66 | |
67 | The first field is a bit mask indicating what cpus this domain operates over. | |
68 | ||
b762f3ff JD |
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 | |
1da177e4 LT |
142 | |
143 | /proc/<pid>/schedstat | |
144 | ---------------- | |
9afa802f | 145 | schedstats also adds a new /proc/<pid>/schedstat file to include some of |
1da177e4 | 146 | the same information on a per-process level. There are three fields in |
b762f3ff JD |
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 | |
1da177e4 LT |
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 | |
b762f3ff | 156 | http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c |