1Version 15 of schedstats dropped counters for some sched_yield:
2yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is
3identical to version 14.
4
5Version 14 of schedstats includes support for sched_domains, which hit the
6mainline kernel in 2.6.20 although it is identical to the stats from version
712 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel
8release).  Some counters make more sense to be per-runqueue; other to be
9per-domain.  Note that domains (and their associated information) will only
10be pertinent and available on machines utilizing CONFIG_SMP.
11
12In version 14 of schedstat, there is at least one level of domain
13statistics for each cpu listed, and there may well be more than one
14domain.  Domains have no particular names in this implementation, but
15the highest numbered one typically arbitrates balancing across all the
16cpus on the machine, while domain0 is the most tightly focused domain,
17sometimes balancing only between pairs of cpus.  At this time, there
18are no architectures which need more than three domain levels. The first
19field in the domain stats is a bit map indicating which cpus are affected
20by that domain.
21
22These fields are counters, and only increment.  Programs which make use
23of these will need to start with a baseline observation and then calculate
24the change in the counters at each subsequent observation.  A perl script
25which does this for many of the fields is available at
26
27    http://eaglet.rain.com/rick/linux/schedstat/
28
29Note that any such script will necessarily be version-specific, as the main
30reason to change versions is changes in the output format.  For those wishing
31to write their own scripts, the fields are described here.
32
33CPU statistics
34--------------
35cpu<N> 1 2 3 4 5 6 7 8 9
36
37First field is a sched_yield() statistic:
38     1) # of times sched_yield() was called
39
40Next three are schedule() statistics:
41     2) This field is a legacy array expiration count field used in the O(1)
42	scheduler. We kept it for ABI compatibility, but it is always set to zero.
43     3) # of times schedule() was called
44     4) # of times schedule() left the processor idle
45
46Next two are try_to_wake_up() statistics:
47     5) # of times try_to_wake_up() was called
48     6) # of times try_to_wake_up() was called to wake up the local cpu
49
50Next three are statistics describing scheduling latency:
51     7) sum of all time spent running by tasks on this processor (in jiffies)
52     8) sum of all time spent waiting to run by tasks on this processor (in
53        jiffies)
54     9) # of timeslices run on this cpu
55
56
57Domain statistics
58-----------------
59One of these is produced per domain for each cpu described. (Note that if
60CONFIG_SMP is not defined, *no* domains are utilized and these lines
61will not appear in the output.)
62
63domain<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
65The first field is a bit mask indicating what cpus this domain operates over.
66
67The next 24 are a variety of load_balance() statistics in grouped into types
68of idleness (idle, busy, and newly idle):
69
70     1) # of times in this domain load_balance() was called when the
71        cpu was idle
72     2) # of times in this domain load_balance() checked but found
73        the load did not require balancing when the cpu was idle
74     3) # of times in this domain load_balance() tried to move one or
75        more tasks and failed, when the cpu was idle
76     4) sum of imbalances discovered (if any) with each call to
77        load_balance() in this domain when the cpu was idle
78     5) # of times in this domain pull_task() was called when the cpu
79        was idle
80     6) # of times in this domain pull_task() was called even though
81        the target task was cache-hot when idle
82     7) # of times in this domain load_balance() was called but did
83        not find a busier queue while the cpu was idle
84     8) # of times in this domain a busier queue was found while the
85        cpu was idle but no busier group was found
86
87     9) # of times in this domain load_balance() was called when the
88        cpu was busy
89    10) # of times in this domain load_balance() checked but found the
90        load did not require balancing when busy
91    11) # of times in this domain load_balance() tried to move one or
92        more tasks and failed, when the cpu was busy
93    12) sum of imbalances discovered (if any) with each call to
94        load_balance() in this domain when the cpu was busy
95    13) # of times in this domain pull_task() was called when busy
96    14) # of times in this domain pull_task() was called even though the
97        target task was cache-hot when busy
98    15) # of times in this domain load_balance() was called but did not
99        find a busier queue while the cpu was busy
100    16) # of times in this domain a busier queue was found while the cpu
101        was busy but no busier group was found
102
103    17) # of times in this domain load_balance() was called when the
104        cpu was just becoming idle
105    18) # of times in this domain load_balance() checked but found the
106        load did not require balancing when the cpu was just becoming idle
107    19) # of times in this domain load_balance() tried to move one or more
108        tasks and failed, when the cpu was just becoming idle
109    20) sum of imbalances discovered (if any) with each call to
110        load_balance() in this domain when the cpu was just becoming idle
111    21) # of times in this domain pull_task() was called when newly idle
112    22) # of times in this domain pull_task() was called even though the
113        target task was cache-hot when just becoming idle
114    23) # of times in this domain load_balance() was called but did not
115        find a busier queue while the cpu was just becoming idle
116    24) # of times in this domain a busier queue was found while the cpu
117        was just becoming idle but no busier group was found
118
119   Next three are active_load_balance() statistics:
120    25) # of times active_load_balance() was called
121    26) # of times active_load_balance() tried to move a task and failed
122    27) # of times active_load_balance() successfully moved a task
123
124   Next three are sched_balance_exec() statistics:
125    28) sbe_cnt is not used
126    29) sbe_balanced is not used
127    30) sbe_pushed is not used
128
129   Next three are sched_balance_fork() statistics:
130    31) sbf_cnt is not used
131    32) sbf_balanced is not used
132    33) sbf_pushed is not used
133
134   Next three are try_to_wake_up() statistics:
135    34) # of times in this domain try_to_wake_up() awoke a task that
136        last ran on a different cpu in this domain
137    35) # of times in this domain try_to_wake_up() moved a task to the
138        waking cpu because it was cache-cold on its own cpu anyway
139    36) # of times in this domain try_to_wake_up() started passive balancing
140
141/proc/<pid>/schedstat
142----------------
143schedstats also adds a new /proc/<pid>/schedstat file to include some of
144the same information on a per-process level.  There are three fields in
145this file correlating for that process to:
146     1) time spent on the cpu
147     2) time spent waiting on a runqueue
148     3) # of timeslices run on this cpu
149
150A program could be easily written to make use of these extra fields to
151report on how well a particular process or set of processes is faring
152under the scheduler's policies.  A simple version of such a program is
153available at
154    http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c
155