Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | Version 10 of schedstats includes support for sched_domains, which |
2 | hit the mainline kernel in 2.6.7. Some counters make more sense to be | |
3 | per-runqueue; other to be per-domain. Note that domains (and their associated | |
4 | information) will only be pertinent and available on machines utilizing | |
5 | CONFIG_SMP. | |
6 | ||
7 | In version 10 of schedstat, there is at least one level of domain | |
8 | statistics for each cpu listed, and there may well be more than one | |
9 | domain. Domains have no particular names in this implementation, but | |
10 | the highest numbered one typically arbitrates balancing across all the | |
11 | cpus on the machine, while domain0 is the most tightly focused domain, | |
12 | sometimes balancing only between pairs of cpus. At this time, there | |
13 | are no architectures which need more than three domain levels. The first | |
14 | field in the domain stats is a bit map indicating which cpus are affected | |
15 | by that domain. | |
16 | ||
17 | These fields are counters, and only increment. Programs which make use | |
18 | of these will need to start with a baseline observation and then calculate | |
19 | the change in the counters at each subsequent observation. A perl script | |
20 | which does this for many of the fields is available at | |
21 | ||
22 | http://eaglet.rain.com/rick/linux/schedstat/ | |
23 | ||
24 | Note that any such script will necessarily be version-specific, as the main | |
25 | reason to change versions is changes in the output format. For those wishing | |
26 | to write their own scripts, the fields are described here. | |
27 | ||
28 | CPU statistics | |
29 | -------------- | |
30 | cpu<N> 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 | |
31 | ||
32 | NOTE: In the sched_yield() statistics, the active queue is considered empty | |
33 | if it has only one process in it, since obviously the process calling | |
34 | sched_yield() is that process. | |
35 | ||
36 | First four fields are sched_yield() statistics: | |
37 | 1) # of times both the active and the expired queue were empty | |
38 | 2) # of times just the active queue was empty | |
39 | 3) # of times just the expired queue was empty | |
40 | 4) # of times sched_yield() was called | |
41 | ||
42 | Next four are schedule() statistics: | |
43 | 5) # of times the active queue had at least one other process on it | |
44 | 6) # of times we switched to the expired queue and reused it | |
45 | 7) # of times schedule() was called | |
46 | 8) # of times schedule() left the processor idle | |
47 | ||
48 | Next four are active_load_balance() statistics: | |
49 | 9) # of times active_load_balance() was called | |
50 | 10) # of times active_load_balance() caused this cpu to gain a task | |
51 | 11) # of times active_load_balance() caused this cpu to lose a task | |
52 | 12) # of times active_load_balance() tried to move a task and failed | |
53 | ||
54 | Next three are try_to_wake_up() statistics: | |
55 | 13) # of times try_to_wake_up() was called | |
56 | 14) # of times try_to_wake_up() successfully moved the awakening task | |
57 | 15) # of times try_to_wake_up() attempted to move the awakening task | |
58 | ||
59 | Next two are wake_up_new_task() statistics: | |
60 | 16) # of times wake_up_new_task() was called | |
61 | 17) # of times wake_up_new_task() successfully moved the new task | |
62 | ||
63 | Next one is a sched_migrate_task() statistic: | |
64 | 18) # of times sched_migrate_task() was called | |
65 | ||
66 | Next one is a sched_balance_exec() statistic: | |
67 | 19) # of times sched_balance_exec() was called | |
68 | ||
69 | Next three are statistics describing scheduling latency: | |
70 | 20) sum of all time spent running by tasks on this processor (in ms) | |
71 | 21) sum of all time spent waiting to run by tasks on this processor (in ms) | |
72 | 22) # of tasks (not necessarily unique) given to the processor | |
73 | ||
74 | The last six are statistics dealing with pull_task(): | |
75 | 23) # of times pull_task() moved a task to this cpu when newly idle | |
76 | 24) # of times pull_task() stole a task from this cpu when another cpu | |
77 | was newly idle | |
78 | 25) # of times pull_task() moved a task to this cpu when idle | |
79 | 26) # of times pull_task() stole a task from this cpu when another cpu | |
80 | was idle | |
81 | 27) # of times pull_task() moved a task to this cpu when busy | |
82 | 28) # of times pull_task() stole a task from this cpu when another cpu | |
83 | was busy | |
84 | ||
85 | ||
86 | Domain statistics | |
87 | ----------------- | |
88 | One of these is produced per domain for each cpu described. (Note that if | |
89 | CONFIG_SMP is not defined, *no* domains are utilized and these lines | |
90 | will not appear in the output.) | |
91 | ||
92 | domain<N> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 | |
93 | ||
94 | The first field is a bit mask indicating what cpus this domain operates over. | |
95 | ||
96 | The next fifteen are a variety of load_balance() statistics: | |
97 | ||
98 | 1) # of times in this domain load_balance() was called when the cpu | |
99 | was idle | |
100 | 2) # of times in this domain load_balance() was called when the cpu | |
101 | was busy | |
102 | 3) # of times in this domain load_balance() was called when the cpu | |
103 | was just becoming idle | |
104 | 4) # of times in this domain load_balance() tried to move one or more | |
105 | tasks and failed, when the cpu was idle | |
106 | 5) # of times in this domain load_balance() tried to move one or more | |
107 | tasks and failed, when the cpu was busy | |
108 | 6) # of times in this domain load_balance() tried to move one or more | |
109 | tasks and failed, when the cpu was just becoming idle | |
110 | 7) sum of imbalances discovered (if any) with each call to | |
111 | load_balance() in this domain when the cpu was idle | |
112 | 8) sum of imbalances discovered (if any) with each call to | |
113 | load_balance() in this domain when the cpu was busy | |
114 | 9) sum of imbalances discovered (if any) with each call to | |
115 | load_balance() in this domain when the cpu was just becoming idle | |
116 | 10) # of times in this domain load_balance() was called but did not find | |
117 | a busier queue while the cpu was idle | |
118 | 11) # of times in this domain load_balance() was called but did not find | |
119 | a busier queue while the cpu was busy | |
120 | 12) # of times in this domain load_balance() was called but did not find | |
121 | a busier queue while the cpu was just becoming idle | |
122 | 13) # of times in this domain a busier queue was found while the cpu was | |
123 | idle but no busier group was found | |
124 | 14) # of times in this domain a busier queue was found while the cpu was | |
125 | busy but no busier group was found | |
126 | 15) # of times in this domain a busier queue was found while the cpu was | |
127 | just becoming idle but no busier group was found | |
128 | ||
129 | Next two are sched_balance_exec() statistics: | |
130 | 17) # of times in this domain sched_balance_exec() successfully pushed | |
131 | a task to a new cpu | |
132 | 18) # of times in this domain sched_balance_exec() tried but failed to | |
133 | push a task to a new cpu | |
134 | ||
135 | Next two are try_to_wake_up() statistics: | |
136 | 19) # of times in this domain try_to_wake_up() tried to move a task based | |
137 | on affinity and cache warmth | |
138 | 20) # of times in this domain try_to_wake_up() tried to move a task based | |
139 | on load balancing | |
140 | ||
141 | ||
142 | /proc/<pid>/schedstat | |
143 | ---------------- | |
144 | schedstats also adds a new /proc/<pid/schedstat file to include some of | |
145 | the same information on a per-process level. There are three fields in | |
146 | this file correlating to fields 20, 21, and 22 in the CPU fields, but | |
147 | they only apply for that process. | |
148 | ||
149 | A program could be easily written to make use of these extra fields to | |
150 | report on how well a particular process or set of processes is faring | |
151 | under the scheduler's policies. A simple version of such a program is | |
152 | available at | |
153 | http://eaglet.rain.com/rick/linux/schedstat/v10/latency.c |