Add the rt linux 4.1.3-rt3 as base
[kvmfornfv.git] / kernel / Documentation / scheduler / sched-stats.txt
1 Version 15 of schedstats dropped counters for some sched_yield:
2 yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is
3 identical to version 14.
4
5 Version 14 of schedstats includes support for sched_domains, which hit the
6 mainline kernel in 2.6.20 although it is identical to the stats from version
7 12 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel
8 release).  Some counters make more sense to be per-runqueue; other to be
9 per-domain.  Note that domains (and their associated information) will only
10 be pertinent and available on machines utilizing CONFIG_SMP.
11
12 In version 14 of schedstat, there is at least one level of domain
13 statistics for each cpu listed, and there may well be more than one
14 domain.  Domains have no particular names in this implementation, but
15 the highest numbered one typically arbitrates balancing across all the
16 cpus on the machine, while domain0 is the most tightly focused domain,
17 sometimes balancing only between pairs of cpus.  At this time, there
18 are no architectures which need more than three domain levels. The first
19 field in the domain stats is a bit map indicating which cpus are affected
20 by that domain.
21
22 These fields are counters, and only increment.  Programs which make use
23 of these will need to start with a baseline observation and then calculate
24 the change in the counters at each subsequent observation.  A perl script
25 which does this for many of the fields is available at
26
27     http://eaglet.rain.com/rick/linux/schedstat/
28
29 Note that any such script will necessarily be version-specific, as the main
30 reason to change versions is changes in the output format.  For those wishing
31 to write their own scripts, the fields are described here.
32
33 CPU statistics
34 --------------
35 cpu<N> 1 2 3 4 5 6 7 8 9
36
37 First field is a sched_yield() statistic:
38      1) # of times sched_yield() was called
39
40 Next 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
46 Next 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
50 Next 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
57 Domain statistics
58 -----------------
59 One of these is produced per domain for each cpu described. (Note that if
60 CONFIG_SMP is not defined, *no* domains are utilized and these lines
61 will not appear in the output.)
62
63 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
64
65 The first field is a bit mask indicating what cpus this domain operates over.
66
67 The next 24 are a variety of load_balance() statistics in grouped into types
68 of 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 ----------------
143 schedstats also adds a new /proc/<pid>/schedstat file to include some of
144 the same information on a per-process level.  There are three fields in
145 this 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
150 A program could be easily written to make use of these extra fields to
151 report on how well a particular process or set of processes is faring
152 under the scheduler's policies.  A simple version of such a program is
153 available at
154     http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c