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