Contents of /alx-src/tags/kernel26-2.6.12-alx-r9/Documentation/sched-stats.txt
Parent Directory | Revision Log
Revision 630 -
(show annotations)
(download)
Wed Mar 4 11:03:09 2009 UTC (15 years, 6 months ago) by niro
File MIME type: text/plain
File size: 7162 byte(s)
Wed Mar 4 11:03:09 2009 UTC (15 years, 6 months ago) by niro
File MIME type: text/plain
File size: 7162 byte(s)
Tag kernel26-2.6.12-alx-r9
1 | Version 10 of schedstats includes support for sched_domains, which |
2 | hit the mainline kernel in 2.6.7. Some counters make more sense to be |
3 | per-runqueue; other to be per-domain. Note that domains (and their associated |
4 | information) will only be pertinent and available on machines utilizing |
5 | CONFIG_SMP. |
6 | |
7 | In version 10 of schedstat, there is at least one level of domain |
8 | statistics for each cpu listed, and there may well be more than one |
9 | domain. Domains have no particular names in this implementation, but |
10 | the highest numbered one typically arbitrates balancing across all the |
11 | cpus on the machine, while domain0 is the most tightly focused domain, |
12 | sometimes balancing only between pairs of cpus. At this time, there |
13 | are no architectures which need more than three domain levels. The first |
14 | field in the domain stats is a bit map indicating which cpus are affected |
15 | by that domain. |
16 | |
17 | These fields are counters, and only increment. Programs which make use |
18 | of these will need to start with a baseline observation and then calculate |
19 | the change in the counters at each subsequent observation. A perl script |
20 | which does this for many of the fields is available at |
21 | |
22 | http://eaglet.rain.com/rick/linux/schedstat/ |
23 | |
24 | Note that any such script will necessarily be version-specific, as the main |
25 | reason to change versions is changes in the output format. For those wishing |
26 | to write their own scripts, the fields are described here. |
27 | |
28 | CPU statistics |
29 | -------------- |
30 | cpu<N> 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 |
31 | |
32 | NOTE: In the sched_yield() statistics, the active queue is considered empty |
33 | if it has only one process in it, since obviously the process calling |
34 | sched_yield() is that process. |
35 | |
36 | First four fields are sched_yield() statistics: |
37 | 1) # of times both the active and the expired queue were empty |
38 | 2) # of times just the active queue was empty |
39 | 3) # of times just the expired queue was empty |
40 | 4) # of times sched_yield() was called |
41 | |
42 | Next four are schedule() statistics: |
43 | 5) # of times the active queue had at least one other process on it |
44 | 6) # of times we switched to the expired queue and reused it |
45 | 7) # of times schedule() was called |
46 | 8) # of times schedule() left the processor idle |
47 | |
48 | Next four are active_load_balance() statistics: |
49 | 9) # of times active_load_balance() was called |
50 | 10) # of times active_load_balance() caused this cpu to gain a task |
51 | 11) # of times active_load_balance() caused this cpu to lose a task |
52 | 12) # of times active_load_balance() tried to move a task and failed |
53 | |
54 | Next three are try_to_wake_up() statistics: |
55 | 13) # of times try_to_wake_up() was called |
56 | 14) # of times try_to_wake_up() successfully moved the awakening task |
57 | 15) # of times try_to_wake_up() attempted to move the awakening task |
58 | |
59 | Next two are wake_up_new_task() statistics: |
60 | 16) # of times wake_up_new_task() was called |
61 | 17) # of times wake_up_new_task() successfully moved the new task |
62 | |
63 | Next one is a sched_migrate_task() statistic: |
64 | 18) # of times sched_migrate_task() was called |
65 | |
66 | Next one is a sched_balance_exec() statistic: |
67 | 19) # of times sched_balance_exec() was called |
68 | |
69 | Next three are statistics describing scheduling latency: |
70 | 20) sum of all time spent running by tasks on this processor (in ms) |
71 | 21) sum of all time spent waiting to run by tasks on this processor (in ms) |
72 | 22) # of tasks (not necessarily unique) given to the processor |
73 | |
74 | The last six are statistics dealing with pull_task(): |
75 | 23) # of times pull_task() moved a task to this cpu when newly idle |
76 | 24) # of times pull_task() stole a task from this cpu when another cpu |
77 | was newly idle |
78 | 25) # of times pull_task() moved a task to this cpu when idle |
79 | 26) # of times pull_task() stole a task from this cpu when another cpu |
80 | was idle |
81 | 27) # of times pull_task() moved a task to this cpu when busy |
82 | 28) # of times pull_task() stole a task from this cpu when another cpu |
83 | was busy |
84 | |
85 | |
86 | Domain statistics |
87 | ----------------- |
88 | One of these is produced per domain for each cpu described. (Note that if |
89 | CONFIG_SMP is not defined, *no* domains are utilized and these lines |
90 | will not appear in the output.) |
91 | |
92 | domain<N> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 |
93 | |
94 | The first field is a bit mask indicating what cpus this domain operates over. |
95 | |
96 | The next fifteen are a variety of load_balance() statistics: |
97 | |
98 | 1) # of times in this domain load_balance() was called when the cpu |
99 | was idle |
100 | 2) # of times in this domain load_balance() was called when the cpu |
101 | was busy |
102 | 3) # of times in this domain load_balance() was called when the cpu |
103 | was just becoming idle |
104 | 4) # of times in this domain load_balance() tried to move one or more |
105 | tasks and failed, when the cpu was idle |
106 | 5) # of times in this domain load_balance() tried to move one or more |
107 | tasks and failed, when the cpu was busy |
108 | 6) # of times in this domain load_balance() tried to move one or more |
109 | tasks and failed, when the cpu was just becoming idle |
110 | 7) sum of imbalances discovered (if any) with each call to |
111 | load_balance() in this domain when the cpu was idle |
112 | 8) sum of imbalances discovered (if any) with each call to |
113 | load_balance() in this domain when the cpu was busy |
114 | 9) sum of imbalances discovered (if any) with each call to |
115 | load_balance() in this domain when the cpu was just becoming idle |
116 | 10) # of times in this domain load_balance() was called but did not find |
117 | a busier queue while the cpu was idle |
118 | 11) # of times in this domain load_balance() was called but did not find |
119 | a busier queue while the cpu was busy |
120 | 12) # of times in this domain load_balance() was called but did not find |
121 | a busier queue while the cpu was just becoming idle |
122 | 13) # of times in this domain a busier queue was found while the cpu was |
123 | idle but no busier group was found |
124 | 14) # of times in this domain a busier queue was found while the cpu was |
125 | busy but no busier group was found |
126 | 15) # of times in this domain a busier queue was found while the cpu was |
127 | just becoming idle but no busier group was found |
128 | |
129 | Next two are sched_balance_exec() statistics: |
130 | 17) # of times in this domain sched_balance_exec() successfully pushed |
131 | a task to a new cpu |
132 | 18) # of times in this domain sched_balance_exec() tried but failed to |
133 | push a task to a new cpu |
134 | |
135 | Next two are try_to_wake_up() statistics: |
136 | 19) # of times in this domain try_to_wake_up() tried to move a task based |
137 | on affinity and cache warmth |
138 | 20) # of times in this domain try_to_wake_up() tried to move a task based |
139 | on load balancing |
140 | |
141 | |
142 | /proc/<pid>/schedstat |
143 | ---------------- |
144 | schedstats also adds a new /proc/<pid/schedstat file to include some of |
145 | the same information on a per-process level. There are three fields in |
146 | this file correlating to fields 20, 21, and 22 in the CPU fields, but |
147 | they only apply for that process. |
148 | |
149 | A program could be easily written to make use of these extra fields to |
150 | report on how well a particular process or set of processes is faring |
151 | under the scheduler's policies. A simple version of such a program is |
152 | available at |
153 | http://eaglet.rain.com/rick/linux/schedstat/v10/latency.c |