Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
1====================
2Scheduler Statistics
3====================
4
5Version 17 of schedstats removed 'lb_imbalance' field as it has no
6significance anymore and instead added more relevant fields namely
7'lb_imbalance_load', 'lb_imbalance_util', 'lb_imbalance_task' and
8'lb_imbalance_misfit'. The domain field prints the name of the
9corresponding sched domain from this version onwards.
10
11Version 16 of schedstats changed the order of definitions within
12'enum cpu_idle_type', which changed the order of [CPU_MAX_IDLE_TYPES]
13columns in show_schedstat(). In particular the position of CPU_IDLE
14and __CPU_NOT_IDLE changed places. The size of the array is unchanged.
15
16Version 15 of schedstats dropped counters for some sched_yield:
17yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is
18identical to version 14. Details are available at
19
20 https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/scheduler/sched-stats.txt?id=1e1dbb259c79b
21
22Version 14 of schedstats includes support for sched_domains, which hit the
23mainline kernel in 2.6.20 although it is identical to the stats from version
2412 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel
25release). Some counters make more sense to be per-runqueue; other to be
26per-domain. Note that domains (and their associated information) will only
27be pertinent and available on machines utilizing CONFIG_SMP.
28
29In version 14 of schedstat, there is at least one level of domain
30statistics for each cpu listed, and there may well be more than one
31domain. Domains have no particular names in this implementation, but
32the highest numbered one typically arbitrates balancing across all the
33cpus on the machine, while domain0 is the most tightly focused domain,
34sometimes balancing only between pairs of cpus. At this time, there
35are no architectures which need more than three domain levels. The first
36field in the domain stats is a bit map indicating which cpus are affected
37by that domain. Details are available at
38
39 https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/sched-stats.txt?id=b762f3ffb797c
40
41The schedstat documentation is maintained version 10 onwards and is not
42updated for version 11 and 12. The details for version 10 are available at
43
44 https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/sched-stats.txt?id=1da177e4c3f4
45
46These fields are counters, and only increment. Programs which make use
47of these will need to start with a baseline observation and then calculate
48the change in the counters at each subsequent observation. A perl script
49which does this for many of the fields is available at
50
51 http://eaglet.pdxhosts.com/rick/linux/schedstat/
52
53Note that any such script will necessarily be version-specific, as the main
54reason to change versions is changes in the output format. For those wishing
55to write their own scripts, the fields are described here.
56
57CPU statistics
58--------------
59cpu<N> 1 2 3 4 5 6 7 8 9
60
61First field is a sched_yield() statistic:
62
63 1) # of times sched_yield() was called
64
65Next three are schedule() statistics:
66
67 2) This field is a legacy array expiration count field used in the O(1)
68 scheduler. We kept it for ABI compatibility, but it is always set to zero.
69 3) # of times schedule() was called
70 4) # of times schedule() left the processor idle
71
72Next two are try_to_wake_up() statistics:
73
74 5) # of times try_to_wake_up() was called
75 6) # of times try_to_wake_up() was called to wake up the local cpu
76
77Next three are statistics describing scheduling latency:
78
79 7) sum of all time spent running by tasks on this processor (in nanoseconds)
80 8) sum of all time spent waiting to run by tasks on this processor (in
81 nanoseconds)
82 9) # of timeslices run on this cpu
83
84
85Domain statistics
86-----------------
87One of these is produced per domain for each cpu described. (Note that if
88CONFIG_SMP is not defined, *no* domains are utilized and these lines
89will not appear in the output.)
90
91domain<N> <name> <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 37 38 39 40 41 42 43 44 45
92
93The <name> field prints the name of the sched domain and is only supported
94with schedstat version >= 17. On previous versions, <cpumask> is the first
95field.
96
97The <cpumask> field is a bit mask indicating what cpus this domain operates
98over.
99
100The next 33 are a variety of sched_balance_rq() statistics in grouped into types
101of idleness (busy, idle and newly idle):
102
103 1) # of times in this domain sched_balance_rq() was called when the
104 cpu was busy
105 2) # of times in this domain sched_balance_rq() checked but found the
106 load did not require balancing when busy
107 3) # of times in this domain sched_balance_rq() tried to move one or
108 more tasks and failed, when the cpu was busy
109 4) Total imbalance in load in this domain when the cpu was busy
110 5) Total imbalance in utilization in this domain when the cpu was busy
111 6) Total imbalance in number of tasks in this domain when the cpu was busy
112 7) Total imbalance due to misfit tasks in this domain when the cpu was
113 busy
114 8) # of times in this domain detach_task() was called when busy
115 9) # of times in this domain detach_task() was called even though the
116 target task was cache-hot when busy
117 10) # of times in this domain sched_balance_rq() was called but did not
118 find a busier queue while the cpu was busy
119 11) # of times in this domain a busier queue was found while the cpu
120 was busy but no busier group was found
121
122 12) # of times in this domain sched_balance_rq() was called when the
123 cpu was idle
124 13) # of times in this domain sched_balance_rq() checked but found
125 the load did not require balancing when the cpu was idle
126 14) # of times in this domain sched_balance_rq() tried to move one or
127 more tasks and failed, when the cpu was idle
128 15) Total imbalance in load in this domain when the cpu was idle
129 16) Total imbalance in utilization in this domain when the cpu was idle
130 17) Total imbalance in number of tasks in this domain when the cpu was idle
131 18) Total imbalance due to misfit tasks in this domain when the cpu was
132 idle
133 19) # of times in this domain detach_task() was called when the cpu
134 was idle
135 20) # of times in this domain detach_task() was called even though
136 the target task was cache-hot when idle
137 21) # of times in this domain sched_balance_rq() was called but did
138 not find a busier queue while the cpu was idle
139 22) # of times in this domain a busier queue was found while the
140 cpu was idle but no busier group was found
141
142 23) # of times in this domain sched_balance_rq() was called when the
143 cpu was just becoming idle
144 24) # of times in this domain sched_balance_rq() checked but found the
145 load did not require balancing when the cpu was just becoming idle
146 25) # of times in this domain sched_balance_rq() tried to move one or more
147 tasks and failed, when the cpu was just becoming idle
148 26) Total imbalance in load in this domain when the cpu was just becoming
149 idle
150 27) Total imbalance in utilization in this domain when the cpu was just
151 becoming idle
152 28) Total imbalance in number of tasks in this domain when the cpu was just
153 becoming idle
154 29) Total imbalance due to misfit tasks in this domain when the cpu was
155 just becoming idle
156 30) # of times in this domain detach_task() was called when newly idle
157 31) # of times in this domain detach_task() was called even though the
158 target task was cache-hot when just becoming idle
159 32) # of times in this domain sched_balance_rq() was called but did not
160 find a busier queue while the cpu was just becoming idle
161 33) # of times in this domain a busier queue was found while the cpu
162 was just becoming idle but no busier group was found
163
164 Next three are active_load_balance() statistics:
165
166 34) # of times active_load_balance() was called
167 35) # of times active_load_balance() tried to move a task and failed
168 36) # of times active_load_balance() successfully moved a task
169
170 Next three are sched_balance_exec() statistics:
171
172 37) sbe_cnt is not used
173 38) sbe_balanced is not used
174 39) sbe_pushed is not used
175
176 Next three are sched_balance_fork() statistics:
177
178 40) sbf_cnt is not used
179 41) sbf_balanced is not used
180 42) sbf_pushed is not used
181
182 Next three are try_to_wake_up() statistics:
183
184 43) # of times in this domain try_to_wake_up() awoke a task that
185 last ran on a different cpu in this domain
186 44) # of times in this domain try_to_wake_up() moved a task to the
187 waking cpu because it was cache-cold on its own cpu anyway
188 45) # of times in this domain try_to_wake_up() started passive balancing
189
190/proc/<pid>/schedstat
191---------------------
192schedstats also adds a new /proc/<pid>/schedstat file to include some of
193the same information on a per-process level. There are three fields in
194this file correlating for that process to:
195
196 1) time spent on the cpu (in nanoseconds)
197 2) time spent waiting on a runqueue (in nanoseconds)
198 3) # of timeslices run on this cpu
199
200A program could be easily written to make use of these extra fields to
201report on how well a particular process or set of processes is faring
202under the scheduler's policies. A simple version of such a program is
203available at
204
205 http://eaglet.pdxhosts.com/rick/linux/schedstat/v12/latency.c