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