1 ====================
2 Scheduler Statistics
3 ====================
4
5 Version 16 of schedstats changed the order of
6 'enum cpu_idle_type', which changed the order
7 columns in show_schedstat(). In particular the
8 and __CPU_NOT_IDLE changed places. The size of
9
10 Version 15 of schedstats dropped counters for
11 yld_exp_empty, yld_act_empty and yld_both_empt
12 identical to version 14.
13
14 Version 14 of schedstats includes support for
15 mainline kernel in 2.6.20 although it is ident
16 12 which was in the kernel from 2.6.13-2.6.19
17 release). Some counters make more sense to be
18 per-domain. Note that domains (and their asso
19 be pertinent and available on machines utilizi
20
21 In version 14 of schedstat, there is at least
22 statistics for each cpu listed, and there may
23 domain. Domains have no particular names in t
24 the highest numbered one typically arbitrates
25 cpus on the machine, while domain0 is the most
26 sometimes balancing only between pairs of cpus
27 are no architectures which need more than thre
28 field in the domain stats is a bit map indicat
29 by that domain.
30
31 These fields are counters, and only increment.
32 of these will need to start with a baseline ob
33 the change in the counters at each subsequent
34 which does this for many of the fields is avai
35
36 http://eaglet.pdxhosts.com/rick/linux/sche
37
38 Note that any such script will necessarily be
39 reason to change versions is changes in the ou
40 to write their own scripts, the fields are des
41
42 CPU statistics
43 --------------
44 cpu<N> 1 2 3 4 5 6 7 8 9
45
46 First field is a sched_yield() statistic:
47
48 1) # of times sched_yield() was called
49
50 Next three are schedule() statistics:
51
52 2) This field is a legacy array expiratio
53 scheduler. We kept it for ABI compatib
54 3) # of times schedule() was called
55 4) # of times schedule() left the process
56
57 Next 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
61
62 Next three are statistics describing schedulin
63
64 7) sum of all time spent running by tasks
65 8) sum of all time spent waiting to run b
66 nanoseconds)
67 9) # of timeslices run on this cpu
68
69
70 Domain statistics
71 -----------------
72 One of these is produced per domain for each c
73 CONFIG_SMP is not defined, *no* domains are ut
74 will not appear in the output.)
75
76 domain<N> <cpumask> 1 2 3 4 5 6 7 8 9 10 11 12
77
78 The first field is a bit mask indicating what
79
80 The next 24 are a variety of sched_balance_rq(
81 of idleness (idle, busy, and newly idle):
82
83 1) # of times in this domain sched_balanc
84 cpu was idle
85 2) # of times in this domain sched_balanc
86 the load did not require balancing whe
87 3) # of times in this domain sched_balanc
88 more tasks and failed, when the cpu wa
89 4) sum of imbalances discovered (if any)
90 sched_balance_rq() in this domain when
91 5) # of times in this domain pull_task()
92 was idle
93 6) # of times in this domain pull_task()
94 the target task was cache-hot when idl
95 7) # of times in this domain sched_balanc
96 not find a busier queue while the cpu
97 8) # of times in this domain a busier que
98 cpu was idle but no busier group was f
99 9) # of times in this domain sched_balanc
100 cpu was busy
101 10) # of times in this domain sched_balanc
102 load did not require balancing when bu
103 11) # of times in this domain sched_balanc
104 more tasks and failed, when the cpu wa
105 12) sum of imbalances discovered (if any)
106 sched_balance_rq() in this domain when
107 13) # of times in this domain pull_task()
108 14) # of times in this domain pull_task()
109 target task was cache-hot when busy
110 15) # of times in this domain sched_balanc
111 find a busier queue while the cpu was
112 16) # of times in this domain a busier que
113 was busy but no busier group was found
114
115 17) # of times in this domain sched_balanc
116 cpu was just becoming idle
117 18) # of times in this domain sched_balanc
118 load did not require balancing when th
119 19) # of times in this domain sched_balanc
120 tasks and failed, when the cpu was jus
121 20) sum of imbalances discovered (if any)
122 sched_balance_rq() in this domain when
123 21) # of times in this domain pull_task()
124 22) # of times in this domain pull_task()
125 target task was cache-hot when just be
126 23) # of times in this domain sched_balanc
127 find a busier queue while the cpu was
128 24) # of times in this domain a busier que
129 was just becoming idle but no busier g
130
131 Next three are active_load_balance() statis
132
133 25) # of times active_load_balance() was c
134 26) # of times active_load_balance() tried
135 27) # of times active_load_balance() succe
136
137 Next three are sched_balance_exec() statist
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() statist
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_
152 last ran on a different cpu in this do
153 35) # of times in this domain try_to_wake_
154 waking cpu because it was cache-cold o
155 36) # of times in this domain try_to_wake_
156
157 /proc/<pid>/schedstat
158 ---------------------
159 schedstats also adds a new /proc/<pid>/schedst
160 the same information on a per-process level.
161 this 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 n
165 3) # of timeslices run on this cpu
166
167 A program could be easily written to make use
168 report on how well a particular process or set
169 under the scheduler's policies. A simple vers
170 available at
171
172 http://eaglet.pdxhosts.com/rick/linux/sche
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.