Loading...
Note: File does not exist in v4.6.
1====================
2Scheduler Statistics
3====================
4
5Version 16 of schedstats changed the order of definitions within
6'enum cpu_idle_type', which changed the order of [CPU_MAX_IDLE_TYPES]
7columns in show_schedstat(). In particular the position of CPU_IDLE
8and __CPU_NOT_IDLE changed places. The size of the array is unchanged.
9
10Version 15 of schedstats dropped counters for some sched_yield:
11yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is
12identical to version 14.
13
14Version 14 of schedstats includes support for sched_domains, which hit the
15mainline kernel in 2.6.20 although it is identical to the stats from version
1612 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel
17release). Some counters make more sense to be per-runqueue; other to be
18per-domain. Note that domains (and their associated information) will only
19be pertinent and available on machines utilizing CONFIG_SMP.
20
21In version 14 of schedstat, there is at least one level of domain
22statistics for each cpu listed, and there may well be more than one
23domain. Domains have no particular names in this implementation, but
24the highest numbered one typically arbitrates balancing across all the
25cpus on the machine, while domain0 is the most tightly focused domain,
26sometimes balancing only between pairs of cpus. At this time, there
27are no architectures which need more than three domain levels. The first
28field in the domain stats is a bit map indicating which cpus are affected
29by that domain.
30
31These fields are counters, and only increment. Programs which make use
32of these will need to start with a baseline observation and then calculate
33the change in the counters at each subsequent observation. A perl script
34which does this for many of the fields is available at
35
36 http://eaglet.pdxhosts.com/rick/linux/schedstat/
37
38Note that any such script will necessarily be version-specific, as the main
39reason to change versions is changes in the output format. For those wishing
40to write their own scripts, the fields are described here.
41
42CPU statistics
43--------------
44cpu<N> 1 2 3 4 5 6 7 8 9
45
46First field is a sched_yield() statistic:
47
48 1) # of times sched_yield() was called
49
50Next three are schedule() statistics:
51
52 2) This field is a legacy array expiration count field used in the O(1)
53 scheduler. We kept it for ABI compatibility, but it is always set to zero.
54 3) # of times schedule() was called
55 4) # of times schedule() left the processor idle
56
57Next two are try_to_wake_up() statistics:
58
59 5) # of times try_to_wake_up() was called
60 6) # of times try_to_wake_up() was called to wake up the local cpu
61
62Next three are statistics describing scheduling latency:
63
64 7) sum of all time spent running by tasks on this processor (in nanoseconds)
65 8) sum of all time spent waiting to run by tasks on this processor (in
66 nanoseconds)
67 9) # of timeslices run on this cpu
68
69
70Domain statistics
71-----------------
72One of these is produced per domain for each cpu described. (Note that if
73CONFIG_SMP is not defined, *no* domains are utilized and these lines
74will not appear in the output.)
75
76domain<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
77
78The first field is a bit mask indicating what cpus this domain operates over.
79
80The next 24 are a variety of sched_balance_rq() statistics in grouped into types
81of idleness (idle, busy, and newly idle):
82
83 1) # of times in this domain sched_balance_rq() was called when the
84 cpu was idle
85 2) # of times in this domain sched_balance_rq() checked but found
86 the load did not require balancing when the cpu was idle
87 3) # of times in this domain sched_balance_rq() tried to move one or
88 more tasks and failed, when the cpu was idle
89 4) sum of imbalances discovered (if any) with each call to
90 sched_balance_rq() in this domain when the cpu was idle
91 5) # of times in this domain pull_task() was called when the cpu
92 was idle
93 6) # of times in this domain pull_task() was called even though
94 the target task was cache-hot when idle
95 7) # of times in this domain sched_balance_rq() was called but did
96 not find a busier queue while the cpu was idle
97 8) # of times in this domain a busier queue was found while the
98 cpu was idle but no busier group was found
99 9) # of times in this domain sched_balance_rq() was called when the
100 cpu was busy
101 10) # of times in this domain sched_balance_rq() checked but found the
102 load did not require balancing when busy
103 11) # of times in this domain sched_balance_rq() tried to move one or
104 more tasks and failed, when the cpu was busy
105 12) sum of imbalances discovered (if any) with each call to
106 sched_balance_rq() in this domain when the cpu was busy
107 13) # of times in this domain pull_task() was called when busy
108 14) # of times in this domain pull_task() was called even though the
109 target task was cache-hot when busy
110 15) # of times in this domain sched_balance_rq() was called but did not
111 find a busier queue while the cpu was busy
112 16) # of times in this domain a busier queue was found while the cpu
113 was busy but no busier group was found
114
115 17) # of times in this domain sched_balance_rq() was called when the
116 cpu was just becoming idle
117 18) # of times in this domain sched_balance_rq() checked but found the
118 load did not require balancing when the cpu was just becoming idle
119 19) # of times in this domain sched_balance_rq() tried to move one or more
120 tasks and failed, when the cpu was just becoming idle
121 20) sum of imbalances discovered (if any) with each call to
122 sched_balance_rq() in this domain when the cpu was just becoming idle
123 21) # of times in this domain pull_task() was called when newly idle
124 22) # of times in this domain pull_task() was called even though the
125 target task was cache-hot when just becoming idle
126 23) # of times in this domain sched_balance_rq() was called but did not
127 find a busier queue while the cpu was just becoming idle
128 24) # of times in this domain a busier queue was found while the cpu
129 was just becoming idle but no busier group was found
130
131 Next three are active_load_balance() statistics:
132
133 25) # of times active_load_balance() was called
134 26) # of times active_load_balance() tried to move a task and failed
135 27) # of times active_load_balance() successfully moved a task
136
137 Next three are sched_balance_exec() statistics:
138
139 28) sbe_cnt is not used
140 29) sbe_balanced is not used
141 30) sbe_pushed is not used
142
143 Next three are sched_balance_fork() statistics:
144
145 31) sbf_cnt is not used
146 32) sbf_balanced is not used
147 33) sbf_pushed is not used
148
149 Next three are try_to_wake_up() statistics:
150
151 34) # of times in this domain try_to_wake_up() awoke a task that
152 last ran on a different cpu in this domain
153 35) # of times in this domain try_to_wake_up() moved a task to the
154 waking cpu because it was cache-cold on its own cpu anyway
155 36) # of times in this domain try_to_wake_up() started passive balancing
156
157/proc/<pid>/schedstat
158---------------------
159schedstats also adds a new /proc/<pid>/schedstat file to include some of
160the same information on a per-process level. There are three fields in
161this file correlating for that process to:
162
163 1) time spent on the cpu (in nanoseconds)
164 2) time spent waiting on a runqueue (in nanoseconds)
165 3) # of timeslices run on this cpu
166
167A program could be easily written to make use of these extra fields to
168report on how well a particular process or set of processes is faring
169under the scheduler's policies. A simple version of such a program is
170available at
171
172 http://eaglet.pdxhosts.com/rick/linux/schedstat/v12/latency.c