Commit | Line | Data |
---|---|---|
d6a3b247 MCC |
1 | ==================== |
2 | Scheduler Statistics | |
3 | ==================== | |
4 | ||
1e1dbb25 JM |
5 | Version 15 of schedstats dropped counters for some sched_yield: |
6 | yld_exp_empty, yld_act_empty and yld_both_empty. Otherwise, it is | |
7 | identical to version 14. | |
8 | ||
b762f3ff JD |
9 | Version 14 of schedstats includes support for sched_domains, which hit the |
10 | mainline kernel in 2.6.20 although it is identical to the stats from version | |
11 | 12 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel | |
12 | release). Some counters make more sense to be per-runqueue; other to be | |
13 | per-domain. Note that domains (and their associated information) will only | |
14 | be pertinent and available on machines utilizing CONFIG_SMP. | |
15 | ||
16 | In version 14 of schedstat, there is at least one level of domain | |
1da177e4 LT |
17 | statistics for each cpu listed, and there may well be more than one |
18 | domain. Domains have no particular names in this implementation, but | |
19 | the highest numbered one typically arbitrates balancing across all the | |
20 | cpus on the machine, while domain0 is the most tightly focused domain, | |
21 | sometimes balancing only between pairs of cpus. At this time, there | |
22 | are no architectures which need more than three domain levels. The first | |
23 | field in the domain stats is a bit map indicating which cpus are affected | |
24 | by that domain. | |
25 | ||
26 | These fields are counters, and only increment. Programs which make use | |
27 | of these will need to start with a baseline observation and then calculate | |
28 | the change in the counters at each subsequent observation. A perl script | |
29 | which does this for many of the fields is available at | |
30 | ||
d41abfd7 | 31 | http://eaglet.pdxhosts.com/rick/linux/schedstat/ |
1da177e4 LT |
32 | |
33 | Note that any such script will necessarily be version-specific, as the main | |
34 | reason to change versions is changes in the output format. For those wishing | |
35 | to write their own scripts, the fields are described here. | |
36 | ||
37 | CPU statistics | |
38 | -------------- | |
1e1dbb25 | 39 | cpu<N> 1 2 3 4 5 6 7 8 9 |
1da177e4 | 40 | |
1e1dbb25 | 41 | First field is a sched_yield() statistic: |
d6a3b247 | 42 | |
1e1dbb25 | 43 | 1) # of times sched_yield() was called |
1da177e4 | 44 | |
b762f3ff | 45 | Next three are schedule() statistics: |
d6a3b247 | 46 | |
30fd049a RM |
47 | 2) This field is a legacy array expiration count field used in the O(1) |
48 | scheduler. We kept it for ABI compatibility, but it is always set to zero. | |
1e1dbb25 JM |
49 | 3) # of times schedule() was called |
50 | 4) # of times schedule() left the processor idle | |
1da177e4 | 51 | |
b762f3ff | 52 | Next two are try_to_wake_up() statistics: |
d6a3b247 | 53 | |
1e1dbb25 JM |
54 | 5) # of times try_to_wake_up() was called |
55 | 6) # of times try_to_wake_up() was called to wake up the local cpu | |
1da177e4 LT |
56 | |
57 | Next three are statistics describing scheduling latency: | |
d6a3b247 | 58 | |
ddb21d27 | 59 | 7) sum of all time spent running by tasks on this processor (in nanoseconds) |
1e1dbb25 | 60 | 8) sum of all time spent waiting to run by tasks on this processor (in |
ddb21d27 | 61 | nanoseconds) |
1e1dbb25 | 62 | 9) # of timeslices run on this cpu |
1da177e4 LT |
63 | |
64 | ||
65 | Domain statistics | |
66 | ----------------- | |
67 | One of these is produced per domain for each cpu described. (Note that if | |
68 | CONFIG_SMP is not defined, *no* domains are utilized and these lines | |
69 | will not appear in the output.) | |
70 | ||
b762f3ff | 71 | 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 |
1da177e4 LT |
72 | |
73 | The first field is a bit mask indicating what cpus this domain operates over. | |
74 | ||
b762f3ff JD |
75 | The next 24 are a variety of load_balance() statistics in grouped into types |
76 | of idleness (idle, busy, and newly idle): | |
77 | ||
d6a3b247 | 78 | 1) # of times in this domain load_balance() was called when the |
b762f3ff | 79 | cpu was idle |
d6a3b247 | 80 | 2) # of times in this domain load_balance() checked but found |
b762f3ff | 81 | the load did not require balancing when the cpu was idle |
d6a3b247 | 82 | 3) # of times in this domain load_balance() tried to move one or |
b762f3ff | 83 | more tasks and failed, when the cpu was idle |
d6a3b247 | 84 | 4) sum of imbalances discovered (if any) with each call to |
b762f3ff | 85 | load_balance() in this domain when the cpu was idle |
d6a3b247 | 86 | 5) # of times in this domain pull_task() was called when the cpu |
b762f3ff | 87 | was idle |
d6a3b247 | 88 | 6) # of times in this domain pull_task() was called even though |
b762f3ff | 89 | the target task was cache-hot when idle |
d6a3b247 | 90 | 7) # of times in this domain load_balance() was called but did |
b762f3ff | 91 | not find a busier queue while the cpu was idle |
d6a3b247 | 92 | 8) # of times in this domain a busier queue was found while the |
b762f3ff | 93 | cpu was idle but no busier group was found |
d6a3b247 | 94 | 9) # of times in this domain load_balance() was called when the |
b762f3ff JD |
95 | cpu was busy |
96 | 10) # of times in this domain load_balance() checked but found the | |
97 | load did not require balancing when busy | |
98 | 11) # of times in this domain load_balance() tried to move one or | |
99 | more tasks and failed, when the cpu was busy | |
100 | 12) sum of imbalances discovered (if any) with each call to | |
101 | load_balance() in this domain when the cpu was busy | |
102 | 13) # of times in this domain pull_task() was called when busy | |
103 | 14) # of times in this domain pull_task() was called even though the | |
104 | target task was cache-hot when busy | |
105 | 15) # of times in this domain load_balance() was called but did not | |
106 | find a busier queue while the cpu was busy | |
107 | 16) # of times in this domain a busier queue was found while the cpu | |
108 | was busy but no busier group was found | |
109 | ||
110 | 17) # of times in this domain load_balance() was called when the | |
111 | cpu was just becoming idle | |
112 | 18) # of times in this domain load_balance() checked but found the | |
113 | load did not require balancing when the cpu was just becoming idle | |
114 | 19) # of times in this domain load_balance() tried to move one or more | |
115 | tasks and failed, when the cpu was just becoming idle | |
116 | 20) sum of imbalances discovered (if any) with each call to | |
117 | load_balance() in this domain when the cpu was just becoming idle | |
118 | 21) # of times in this domain pull_task() was called when newly idle | |
119 | 22) # of times in this domain pull_task() was called even though the | |
120 | target task was cache-hot when just becoming idle | |
121 | 23) # of times in this domain load_balance() was called but did not | |
122 | find a busier queue while the cpu was just becoming idle | |
123 | 24) # of times in this domain a busier queue was found while the cpu | |
124 | was just becoming idle but no busier group was found | |
125 | ||
126 | Next three are active_load_balance() statistics: | |
d6a3b247 | 127 | |
b762f3ff JD |
128 | 25) # of times active_load_balance() was called |
129 | 26) # of times active_load_balance() tried to move a task and failed | |
130 | 27) # of times active_load_balance() successfully moved a task | |
131 | ||
132 | Next three are sched_balance_exec() statistics: | |
d6a3b247 | 133 | |
b762f3ff JD |
134 | 28) sbe_cnt is not used |
135 | 29) sbe_balanced is not used | |
136 | 30) sbe_pushed is not used | |
137 | ||
138 | Next three are sched_balance_fork() statistics: | |
d6a3b247 | 139 | |
b762f3ff JD |
140 | 31) sbf_cnt is not used |
141 | 32) sbf_balanced is not used | |
142 | 33) sbf_pushed is not used | |
143 | ||
144 | Next three are try_to_wake_up() statistics: | |
d6a3b247 | 145 | |
b762f3ff JD |
146 | 34) # of times in this domain try_to_wake_up() awoke a task that |
147 | last ran on a different cpu in this domain | |
148 | 35) # of times in this domain try_to_wake_up() moved a task to the | |
149 | waking cpu because it was cache-cold on its own cpu anyway | |
150 | 36) # of times in this domain try_to_wake_up() started passive balancing | |
1da177e4 LT |
151 | |
152 | /proc/<pid>/schedstat | |
d6a3b247 | 153 | --------------------- |
9afa802f | 154 | schedstats also adds a new /proc/<pid>/schedstat file to include some of |
1da177e4 | 155 | the same information on a per-process level. There are three fields in |
b762f3ff | 156 | this file correlating for that process to: |
d6a3b247 | 157 | |
6c757e9f JTH |
158 | 1) time spent on the cpu (in nanoseconds) |
159 | 2) time spent waiting on a runqueue (in nanoseconds) | |
b762f3ff | 160 | 3) # of timeslices run on this cpu |
1da177e4 LT |
161 | |
162 | A program could be easily written to make use of these extra fields to | |
163 | report on how well a particular process or set of processes is faring | |
164 | under the scheduler's policies. A simple version of such a program is | |
165 | available at | |
d6a3b247 | 166 | |
d41abfd7 | 167 | http://eaglet.pdxhosts.com/rick/linux/schedstat/v12/latency.c |