1====================
2Scheduler Statistics
3====================
4
5Version 15 of schedstats dropped counters for some sched_yield:
6yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is
7identical to version 14.
8
9Version 14 of schedstats includes support for sched_domains, which hit the
10mainline kernel in 2.6.20 although it is identical to the stats from version
1112 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel
12release).  Some counters make more sense to be per-runqueue; other to be
13per-domain.  Note that domains (and their associated information) will only
14be pertinent and available on machines utilizing CONFIG_SMP.
15
16In version 14 of schedstat, there is at least one level of domain
17statistics for each cpu listed, and there may well be more than one
18domain.  Domains have no particular names in this implementation, but
19the highest numbered one typically arbitrates balancing across all the
20cpus on the machine, while domain0 is the most tightly focused domain,
21sometimes balancing only between pairs of cpus.  At this time, there
22are no architectures which need more than three domain levels. The first
23field in the domain stats is a bit map indicating which cpus are affected
24by that domain.
25
26These fields are counters, and only increment.  Programs which make use
27of these will need to start with a baseline observation and then calculate
28the change in the counters at each subsequent observation.  A perl script
29which does this for many of the fields is available at
30
31    http://eaglet.pdxhosts.com/rick/linux/schedstat/
32
33Note that any such script will necessarily be version-specific, as the main
34reason to change versions is changes in the output format.  For those wishing
35to write their own scripts, the fields are described here.
36
37CPU statistics
38--------------
39cpu<N> 1 2 3 4 5 6 7 8 9
40
41First field is a sched_yield() statistic:
42
43     1) # of times sched_yield() was called
44
45Next three are schedule() statistics:
46
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
51
52Next two are try_to_wake_up() statistics:
53
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
56
57Next three are statistics describing scheduling latency:
58
59     7) sum of all time spent running by tasks on this processor (in nanoseconds)
60     8) sum of all time spent waiting to run by tasks on this processor (in
61        nanoseconds)
62     9) # of timeslices run on this cpu
63
64
65Domain statistics
66-----------------
67One of these is produced per domain for each cpu described. (Note that if
68CONFIG_SMP is not defined, *no* domains are utilized and these lines
69will not appear in the output.)
70
71domain<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
72
73The first field is a bit mask indicating what cpus this domain operates over.
74
75The next 24 are a variety of load_balance() statistics in grouped into types
76of idleness (idle, busy, and newly idle):
77
78    1)  # of times in this domain load_balance() was called when the
79        cpu was idle
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
87        was idle
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
95        cpu was busy
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
109
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
125
126   Next three are active_load_balance() statistics:
127
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
131
132   Next three are sched_balance_exec() statistics:
133
134    28) sbe_cnt is not used
135    29) sbe_balanced is not used
136    30) sbe_pushed is not used
137
138   Next three are sched_balance_fork() statistics:
139
140    31) sbf_cnt is not used
141    32) sbf_balanced is not used
142    33) sbf_pushed is not used
143
144   Next three are try_to_wake_up() statistics:
145
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
151
152/proc/<pid>/schedstat
153---------------------
154schedstats also adds a new /proc/<pid>/schedstat file to include some of
155the same information on a per-process level.  There are three fields in
156this file correlating for that process to:
157
158     1) time spent on the cpu (in nanoseconds)
159     2) time spent waiting on a runqueue (in nanoseconds)
160     3) # of timeslices run on this cpu
161
162A program could be easily written to make use of these extra fields to
163report on how well a particular process or set of processes is faring
164under the scheduler's policies.  A simple version of such a program is
165available at
166
167    http://eaglet.pdxhosts.com/rick/linux/schedstat/v12/latency.c
168