Root/Documentation/scheduler/sched-stats.txt

1Version 14 of schedstats includes support for sched_domains, which hit the
2mainline kernel in 2.6.20 although it is identical to the stats from version
312 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel
4release). Some counters make more sense to be per-runqueue; other to be
5per-domain. Note that domains (and their associated information) will only
6be pertinent and available on machines utilizing CONFIG_SMP.
7
8In version 14 of schedstat, there is at least one level of domain
9statistics for each cpu listed, and there may well be more than one
10domain. Domains have no particular names in this implementation, but
11the highest numbered one typically arbitrates balancing across all the
12cpus on the machine, while domain0 is the most tightly focused domain,
13sometimes balancing only between pairs of cpus. At this time, there
14are no architectures which need more than three domain levels. The first
15field in the domain stats is a bit map indicating which cpus are affected
16by that domain.
17
18These fields are counters, and only increment. Programs which make use
19of these will need to start with a baseline observation and then calculate
20the change in the counters at each subsequent observation. A perl script
21which does this for many of the fields is available at
22
23    http://eaglet.rain.com/rick/linux/schedstat/
24
25Note that any such script will necessarily be version-specific, as the main
26reason to change versions is changes in the output format. For those wishing
27to write their own scripts, the fields are described here.
28
29CPU statistics
30--------------
31cpu<N> 1 2 3 4 5 6 7 8 9 10 11 12
32
33NOTE: In the sched_yield() statistics, the active queue is considered empty
34    if it has only one process in it, since obviously the process calling
35    sched_yield() is that process.
36
37First four fields are sched_yield() statistics:
38     1) # of times both the active and the expired queue were empty
39     2) # of times just the active queue was empty
40     3) # of times just the expired queue was empty
41     4) # of times sched_yield() was called
42
43Next three are schedule() statistics:
44     5) # of times we switched to the expired queue and reused it
45     6) # of times schedule() was called
46     7) # of times schedule() left the processor idle
47
48Next two are try_to_wake_up() statistics:
49     8) # of times try_to_wake_up() was called
50     9) # of times try_to_wake_up() was called to wake up the local cpu
51
52Next three are statistics describing scheduling latency:
53    10) sum of all time spent running by tasks on this processor (in jiffies)
54    11) sum of all time spent waiting to run by tasks on this processor (in
55        jiffies)
56    12) # of timeslices run on this cpu
57
58
59Domain statistics
60-----------------
61One of these is produced per domain for each cpu described. (Note that if
62CONFIG_SMP is not defined, *no* domains are utilized and these lines
63will not appear in the output.)
64
65domain<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
66
67The first field is a bit mask indicating what cpus this domain operates over.
68
69The next 24 are a variety of load_balance() statistics in grouped into types
70of idleness (idle, busy, and newly idle):
71
72     1) # of times in this domain load_balance() was called when the
73        cpu was idle
74     2) # of times in this domain load_balance() checked but found
75        the load did not require balancing when the cpu was idle
76     3) # of times in this domain load_balance() tried to move one or
77        more tasks and failed, when the cpu was idle
78     4) sum of imbalances discovered (if any) with each call to
79        load_balance() in this domain when the cpu was idle
80     5) # of times in this domain pull_task() was called when the cpu
81        was idle
82     6) # of times in this domain pull_task() was called even though
83        the target task was cache-hot when idle
84     7) # of times in this domain load_balance() was called but did
85        not find a busier queue while the cpu was idle
86     8) # of times in this domain a busier queue was found while the
87        cpu was idle but no busier group was found
88
89     9) # of times in this domain load_balance() was called when the
90        cpu was busy
91    10) # of times in this domain load_balance() checked but found the
92        load did not require balancing when busy
93    11) # of times in this domain load_balance() tried to move one or
94        more tasks and failed, when the cpu was busy
95    12) sum of imbalances discovered (if any) with each call to
96        load_balance() in this domain when the cpu was busy
97    13) # of times in this domain pull_task() was called when busy
98    14) # of times in this domain pull_task() was called even though the
99        target task was cache-hot when busy
100    15) # of times in this domain load_balance() was called but did not
101        find a busier queue while the cpu was busy
102    16) # of times in this domain a busier queue was found while the cpu
103        was busy but no busier group was found
104
105    17) # of times in this domain load_balance() was called when the
106        cpu was just becoming idle
107    18) # of times in this domain load_balance() checked but found the
108        load did not require balancing when the cpu was just becoming idle
109    19) # of times in this domain load_balance() tried to move one or more
110        tasks and failed, when the cpu was just becoming idle
111    20) sum of imbalances discovered (if any) with each call to
112        load_balance() in this domain when the cpu was just becoming idle
113    21) # of times in this domain pull_task() was called when newly idle
114    22) # of times in this domain pull_task() was called even though the
115        target task was cache-hot when just becoming idle
116    23) # of times in this domain load_balance() was called but did not
117        find a busier queue while the cpu was just becoming idle
118    24) # of times in this domain a busier queue was found while the cpu
119        was just becoming idle but no busier group was found
120
121   Next three are active_load_balance() statistics:
122    25) # of times active_load_balance() was called
123    26) # of times active_load_balance() tried to move a task and failed
124    27) # of times active_load_balance() successfully moved a task
125
126   Next three are sched_balance_exec() statistics:
127    28) sbe_cnt is not used
128    29) sbe_balanced is not used
129    30) sbe_pushed is not used
130
131   Next three are sched_balance_fork() statistics:
132    31) sbf_cnt is not used
133    32) sbf_balanced is not used
134    33) sbf_pushed is not used
135
136   Next three are try_to_wake_up() statistics:
137    34) # of times in this domain try_to_wake_up() awoke a task that
138        last ran on a different cpu in this domain
139    35) # of times in this domain try_to_wake_up() moved a task to the
140        waking cpu because it was cache-cold on its own cpu anyway
141    36) # of times in this domain try_to_wake_up() started passive balancing
142
143/proc/<pid>/schedstat
144----------------
145schedstats also adds a new /proc/<pid>/schedstat file to include some of
146the same information on a per-process level. There are three fields in
147this file correlating for that process to:
148     1) time spent on the cpu
149     2) time spent waiting on a runqueue
150     3) # of timeslices run on this cpu
151
152A program could be easily written to make use of these extra fields to
153report on how well a particular process or set of processes is faring
154under the scheduler's policies. A simple version of such a program is
155available at
156    http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c
157

Archive Download this file



interactive