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