aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
-rw-r--r--Documentation/sched-stats.txt195
1 files changed, 99 insertions, 96 deletions
diff --git a/Documentation/sched-stats.txt b/Documentation/sched-stats.txt
index 6f72021aae51..442e14d35dea 100644
--- a/Documentation/sched-stats.txt
+++ b/Documentation/sched-stats.txt
@@ -1,10 +1,11 @@
1Version 10 of schedstats includes support for sched_domains, which 1Version 14 of schedstats includes support for sched_domains, which hit the
2hit the mainline kernel in 2.6.7. Some counters make more sense to be 2mainline kernel in 2.6.20 although it is identical to the stats from version
3per-runqueue; other to be per-domain. Note that domains (and their associated 312 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel
4information) will only be pertinent and available on machines utilizing 4release). Some counters make more sense to be per-runqueue; other to be
5CONFIG_SMP. 5per-domain. Note that domains (and their associated information) will only
6 6be pertinent and available on machines utilizing CONFIG_SMP.
7In version 10 of schedstat, there is at least one level of domain 7
8In version 14 of schedstat, there is at least one level of domain
8statistics for each cpu listed, and there may well be more than one 9statistics for each cpu listed, and there may well be more than one
9domain. Domains have no particular names in this implementation, but 10domain. Domains have no particular names in this implementation, but
10the highest numbered one typically arbitrates balancing across all the 11the highest numbered one typically arbitrates balancing across all the
@@ -27,7 +28,7 @@ to write their own scripts, the fields are described here.
27 28
28CPU statistics 29CPU statistics
29-------------- 30--------------
30cpu<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 31cpu<N> 1 2 3 4 5 6 7 8 9 10 11 12
31 32
32NOTE: In the sched_yield() statistics, the active queue is considered empty 33NOTE: 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 if it has only one process in it, since obviously the process calling
@@ -39,48 +40,20 @@ First four fields are sched_yield() statistics:
39 3) # of times just the expired queue was empty 40 3) # of times just the expired queue was empty
40 4) # of times sched_yield() was called 41 4) # of times sched_yield() was called
41 42
42Next four are schedule() statistics: 43Next three are schedule() statistics:
43 5) # of times the active queue had at least one other process on it 44 5) # of times we switched to the expired queue and reused it
44 6) # of times we switched to the expired queue and reused it 45 6) # of times schedule() was called
45 7) # of times schedule() was called 46 7) # of times schedule() left the processor idle
46 8) # of times schedule() left the processor idle
47
48Next 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
54Next 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
59Next 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
63Next one is a sched_migrate_task() statistic:
64 18) # of times sched_migrate_task() was called
65 47
66Next one is a sched_balance_exec() statistic: 48Next two are try_to_wake_up() statistics:
67 19) # of times sched_balance_exec() was called 49 8) # of times try_to_wake_up() was called
50 9) # of times try_to_wake_up() was called to wake up the local cpu
68 51
69Next three are statistics describing scheduling latency: 52Next three are statistics describing scheduling latency:
70 20) sum of all time spent running by tasks on this processor (in ms) 53 10) sum of all time spent running by tasks on this processor (in jiffies)
71 21) sum of all time spent waiting to run by tasks on this processor (in ms) 54 11) sum of all time spent waiting to run by tasks on this processor (in
72 22) # of tasks (not necessarily unique) given to the processor 55 jiffies)
73 56 12) # of timeslices run on this cpu
74The 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 57
85 58
86Domain statistics 59Domain statistics
@@ -89,65 +62,95 @@ One of these is produced per domain for each cpu described. (Note that if
89CONFIG_SMP is not defined, *no* domains are utilized and these lines 62CONFIG_SMP is not defined, *no* domains are utilized and these lines
90will not appear in the output.) 63will not appear in the output.)
91 64
92domain<N> 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 65domain<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
93 66
94The first field is a bit mask indicating what cpus this domain operates over. 67The first field is a bit mask indicating what cpus this domain operates over.
95 68
96The next fifteen are a variety of load_balance() statistics: 69The next 24 are a variety of load_balance() statistics in grouped into types
97 70of idleness (idle, busy, and newly idle):
98 1) # of times in this domain load_balance() was called when the cpu 71
99 was idle 72 1) # of times in this domain load_balance() was called when the
100 2) # of times in this domain load_balance() was called when the cpu 73 cpu was idle
101 was busy 74 2) # of times in this domain load_balance() checked but found
102 3) # of times in this domain load_balance() was called when the cpu 75 the load did not require balancing when the cpu was idle
103 was just becoming idle 76 3) # of times in this domain load_balance() tried to move one or
104 4) # of times in this domain load_balance() tried to move one or more 77 more tasks and failed, when the cpu was idle
105 tasks and failed, when the cpu was idle 78 4) sum of imbalances discovered (if any) with each call to
106 5) # of times in this domain load_balance() tried to move one or more 79 load_balance() in this domain when the cpu was idle
107 tasks and failed, when the cpu was busy 80 5) # of times in this domain pull_task() was called when the cpu
108 6) # of times in this domain load_balance() tried to move one or more 81 was idle
109 tasks and failed, when the cpu was just becoming idle 82 6) # of times in this domain pull_task() was called even though
110 7) sum of imbalances discovered (if any) with each call to 83 the target task was cache-hot when idle
111 load_balance() in this domain when the cpu was idle 84 7) # of times in this domain load_balance() was called but did
112 8) sum of imbalances discovered (if any) with each call to 85 not find a busier queue while the cpu was idle
113 load_balance() in this domain when the cpu was busy 86 8) # of times in this domain a busier queue was found while the
114 9) sum of imbalances discovered (if any) with each call to 87 cpu was idle but no busier group was found
115 load_balance() in this domain when the cpu was just becoming idle 88
116 10) # of times in this domain load_balance() was called but did not find 89 9) # of times in this domain load_balance() was called when the
117 a busier queue while the cpu was idle 90 cpu was busy
118 11) # of times in this domain load_balance() was called but did not find 91 10) # of times in this domain load_balance() checked but found the
119 a busier queue while the cpu was busy 92 load did not require balancing when busy
120 12) # of times in this domain load_balance() was called but did not find 93 11) # of times in this domain load_balance() tried to move one or
121 a busier queue while the cpu was just becoming idle 94 more tasks and failed, when the cpu was busy
122 13) # of times in this domain a busier queue was found while the cpu was 95 12) sum of imbalances discovered (if any) with each call to
123 idle but no busier group was found 96 load_balance() in this domain when the cpu was busy
124 14) # of times in this domain a busier queue was found while the cpu was 97 13) # of times in this domain pull_task() was called when busy
125 busy but no busier group was found 98 14) # of times in this domain pull_task() was called even though the
126 15) # of times in this domain a busier queue was found while the cpu was 99 target task was cache-hot when busy
127 just becoming idle but no busier group was found 100 15) # of times in this domain load_balance() was called but did not
128 101 find a busier queue while the cpu was busy
129Next two are sched_balance_exec() statistics: 102 16) # of times in this domain a busier queue was found while the cpu
130 17) # of times in this domain sched_balance_exec() successfully pushed 103 was busy but no busier group was found
131 a task to a new cpu 104
132 18) # of times in this domain sched_balance_exec() tried but failed to 105 17) # of times in this domain load_balance() was called when the
133 push a task to a new cpu 106 cpu was just becoming idle
134 107 18) # of times in this domain load_balance() checked but found the
135Next two are try_to_wake_up() statistics: 108 load did not require balancing when the cpu was just becoming idle
136 19) # of times in this domain try_to_wake_up() tried to move a task based 109 19) # of times in this domain load_balance() tried to move one or more
137 on affinity and cache warmth 110 tasks and failed, when the cpu was just becoming idle
138 20) # of times in this domain try_to_wake_up() tried to move a task based 111 20) sum of imbalances discovered (if any) with each call to
139 on load balancing 112 load_balance() in this domain when the cpu was just becoming idle
140 113 21) # of times in this domain pull_task() was called when newly idle
114 22) # of times in this domain pull_task() was called even though the
115 target task was cache-hot when just becoming idle
116 23) # of times in this domain load_balance() was called but did not
117 find a busier queue while the cpu was just becoming idle
118 24) # of times in this domain a busier queue was found while the cpu
119 was just becoming idle but no busier group was found
120
121 Next three are active_load_balance() statistics:
122 25) # of times active_load_balance() was called
123 26) # of times active_load_balance() tried to move a task and failed
124 27) # of times active_load_balance() successfully moved a task
125
126 Next three are sched_balance_exec() statistics:
127 28) sbe_cnt is not used
128 29) sbe_balanced is not used
129 30) sbe_pushed is not used
130
131 Next three are sched_balance_fork() statistics:
132 31) sbf_cnt is not used
133 32) sbf_balanced is not used
134 33) sbf_pushed is not used
135
136 Next three are try_to_wake_up() statistics:
137 34) # of times in this domain try_to_wake_up() awoke a task that
138 last ran on a different cpu in this domain
139 35) # of times in this domain try_to_wake_up() moved a task to the
140 waking cpu because it was cache-cold on its own cpu anyway
141 36) # of times in this domain try_to_wake_up() started passive balancing
141 142
142/proc/<pid>/schedstat 143/proc/<pid>/schedstat
143---------------- 144----------------
144schedstats also adds a new /proc/<pid/schedstat file to include some of 145schedstats also adds a new /proc/<pid/schedstat file to include some of
145the same information on a per-process level. There are three fields in 146the same information on a per-process level. There are three fields in
146this file correlating to fields 20, 21, and 22 in the CPU fields, but 147this file correlating for that process to:
147they only apply for that process. 148 1) time spent on the cpu
149 2) time spent waiting on a runqueue
150 3) # of timeslices run on this cpu
148 151
149A program could be easily written to make use of these extra fields to 152A program could be easily written to make use of these extra fields to
150report on how well a particular process or set of processes is faring 153report on how well a particular process or set of processes is faring
151under the scheduler's policies. A simple version of such a program is 154under the scheduler's policies. A simple version of such a program is
152available at 155available at
153 http://eaglet.rain.com/rick/linux/schedstat/v10/latency.c 156 http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c