blob: 6f72021aae51f5e1a015468f342b2b3355315e3b [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001Version 10 of schedstats includes support for sched_domains, which
2hit the mainline kernel in 2.6.7. Some counters make more sense to be
3per-runqueue; other to be per-domain. Note that domains (and their associated
4information) will only be pertinent and available on machines utilizing
5CONFIG_SMP.
6
7In version 10 of schedstat, there is at least one level of domain
8statistics for each cpu listed, and there may well be more than one
9domain. Domains have no particular names in this implementation, but
10the highest numbered one typically arbitrates balancing across all the
11cpus on the machine, while domain0 is the most tightly focused domain,
12sometimes balancing only between pairs of cpus. At this time, there
13are no architectures which need more than three domain levels. The first
14field in the domain stats is a bit map indicating which cpus are affected
15by that domain.
16
17These fields are counters, and only increment. Programs which make use
18of these will need to start with a baseline observation and then calculate
19the change in the counters at each subsequent observation. A perl script
20which does this for many of the fields is available at
21
22 http://eaglet.rain.com/rick/linux/schedstat/
23
24Note that any such script will necessarily be version-specific, as the main
25reason to change versions is changes in the output format. For those wishing
26to write their own scripts, the fields are described here.
27
28CPU statistics
29--------------
30cpu<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
32NOTE: 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
36First 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
42Next 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
48Next 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
54Next 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
59Next 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
63Next one is a sched_migrate_task() statistic:
64 18) # of times sched_migrate_task() was called
65
66Next one is a sched_balance_exec() statistic:
67 19) # of times sched_balance_exec() was called
68
69Next 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
74The 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
86Domain statistics
87-----------------
88One of these is produced per domain for each cpu described. (Note that if
89CONFIG_SMP is not defined, *no* domains are utilized and these lines
90will not appear in the output.)
91
92domain<N> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
93
94The first field is a bit mask indicating what cpus this domain operates over.
95
96The 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
129Next 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
135Next 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----------------
144schedstats also adds a new /proc/<pid/schedstat file to include some of
145the same information on a per-process level. There are three fields in
146this file correlating to fields 20, 21, and 22 in the CPU fields, but
147they only apply for that process.
148
149A program could be easily written to make use of these extra fields to
150report on how well a particular process or set of processes is faring
151under the scheduler's policies. A simple version of such a program is
152available at
153 http://eaglet.rain.com/rick/linux/schedstat/v10/latency.c