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) # 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
44
45Next 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
48
49Next 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
52        jiffies)
53     9) # of timeslices run on this cpu
54
55
56Domain statistics
57-----------------
58One of these is produced per domain for each cpu described. (Note that if
59CONFIG_SMP is not defined, *no* domains are utilized and these lines
60will not appear in the output.)
61
62domain<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
63
64The first field is a bit mask indicating what cpus this domain operates over.
65
66The next 24 are a variety of load_balance() statistics in grouped into types
67of idleness (idle, busy, and newly idle):
68
69     1) # of times in this domain load_balance() was called when the
70        cpu was idle
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
78        was idle
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
85
86     9) # of times in this domain load_balance() was called when the
87        cpu was busy
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
101
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
117
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
122
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
127
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
132
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
139
140/proc/<pid>/schedstat
141----------------
142schedstats also adds a new /proc/<pid>/schedstat file to include some of
143the same information on a per-process level.  There are three fields in
144this 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
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/v12/latency.c
154