1 |
3 |
xianfeng |
Version 14 of schedstats includes support for sched_domains, which hit the
|
2 |
|
|
mainline kernel in 2.6.20 although it is identical to the stats from version
|
3 |
|
|
12 which was in the kernel from 2.6.13-2.6.19 (version 13 never saw a kernel
|
4 |
|
|
release). Some counters make more sense to be per-runqueue; other to be
|
5 |
|
|
per-domain. Note that domains (and their associated information) will only
|
6 |
|
|
be pertinent and available on machines utilizing CONFIG_SMP.
|
7 |
|
|
|
8 |
|
|
In version 14 of schedstat, there is at least one level of domain
|
9 |
|
|
statistics for each cpu listed, and there may well be more than one
|
10 |
|
|
domain. Domains have no particular names in this implementation, but
|
11 |
|
|
the highest numbered one typically arbitrates balancing across all the
|
12 |
|
|
cpus on the machine, while domain0 is the most tightly focused domain,
|
13 |
|
|
sometimes balancing only between pairs of cpus. At this time, there
|
14 |
|
|
are no architectures which need more than three domain levels. The first
|
15 |
|
|
field in the domain stats is a bit map indicating which cpus are affected
|
16 |
|
|
by that domain.
|
17 |
|
|
|
18 |
|
|
These fields are counters, and only increment. Programs which make use
|
19 |
|
|
of these will need to start with a baseline observation and then calculate
|
20 |
|
|
the change in the counters at each subsequent observation. A perl script
|
21 |
|
|
which does this for many of the fields is available at
|
22 |
|
|
|
23 |
|
|
http://eaglet.rain.com/rick/linux/schedstat/
|
24 |
|
|
|
25 |
|
|
Note that any such script will necessarily be version-specific, as the main
|
26 |
|
|
reason to change versions is changes in the output format. For those wishing
|
27 |
|
|
to write their own scripts, the fields are described here.
|
28 |
|
|
|
29 |
|
|
CPU statistics
|
30 |
|
|
--------------
|
31 |
|
|
cpu 1 2 3 4 5 6 7 8 9 10 11 12
|
32 |
|
|
|
33 |
|
|
NOTE: In the sched_yield() statistics, the active queue is considered empty
|
34 |
|
|
if it has only one process in it, since obviously the process calling
|
35 |
|
|
sched_yield() is that process.
|
36 |
|
|
|
37 |
|
|
First four fields are sched_yield() statistics:
|
38 |
|
|
1) # of times both the active and the expired queue were empty
|
39 |
|
|
2) # of times just the active queue was empty
|
40 |
|
|
3) # of times just the expired queue was empty
|
41 |
|
|
4) # of times sched_yield() was called
|
42 |
|
|
|
43 |
|
|
Next three are schedule() statistics:
|
44 |
|
|
5) # of times we switched to the expired queue and reused it
|
45 |
|
|
6) # of times schedule() was called
|
46 |
|
|
7) # of times schedule() left the processor idle
|
47 |
|
|
|
48 |
|
|
Next two are try_to_wake_up() statistics:
|
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
|
51 |
|
|
|
52 |
|
|
Next three are statistics describing scheduling latency:
|
53 |
|
|
10) sum of all time spent running by tasks on this processor (in jiffies)
|
54 |
|
|
11) sum of all time spent waiting to run by tasks on this processor (in
|
55 |
|
|
jiffies)
|
56 |
|
|
12) # of timeslices run on this cpu
|
57 |
|
|
|
58 |
|
|
|
59 |
|
|
Domain statistics
|
60 |
|
|
-----------------
|
61 |
|
|
One of these is produced per domain for each cpu described. (Note that if
|
62 |
|
|
CONFIG_SMP is not defined, *no* domains are utilized and these lines
|
63 |
|
|
will not appear in the output.)
|
64 |
|
|
|
65 |
|
|
domain 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
|
66 |
|
|
|
67 |
|
|
The first field is a bit mask indicating what cpus this domain operates over.
|
68 |
|
|
|
69 |
|
|
The next 24 are a variety of load_balance() statistics in grouped into types
|
70 |
|
|
of idleness (idle, busy, and newly idle):
|
71 |
|
|
|
72 |
|
|
1) # of times in this domain load_balance() was called when the
|
73 |
|
|
cpu was idle
|
74 |
|
|
2) # of times in this domain load_balance() checked but found
|
75 |
|
|
the load did not require balancing when the cpu was idle
|
76 |
|
|
3) # of times in this domain load_balance() tried to move one or
|
77 |
|
|
more tasks and failed, when the cpu was idle
|
78 |
|
|
4) sum of imbalances discovered (if any) with each call to
|
79 |
|
|
load_balance() in this domain when the cpu was idle
|
80 |
|
|
5) # of times in this domain pull_task() was called when the cpu
|
81 |
|
|
was idle
|
82 |
|
|
6) # of times in this domain pull_task() was called even though
|
83 |
|
|
the target task was cache-hot when idle
|
84 |
|
|
7) # of times in this domain load_balance() was called but did
|
85 |
|
|
not find a busier queue while the cpu was idle
|
86 |
|
|
8) # of times in this domain a busier queue was found while the
|
87 |
|
|
cpu was idle but no busier group was found
|
88 |
|
|
|
89 |
|
|
9) # of times in this domain load_balance() was called when the
|
90 |
|
|
cpu was busy
|
91 |
|
|
10) # of times in this domain load_balance() checked but found the
|
92 |
|
|
load did not require balancing when busy
|
93 |
|
|
11) # of times in this domain load_balance() tried to move one or
|
94 |
|
|
more tasks and failed, when the cpu was busy
|
95 |
|
|
12) sum of imbalances discovered (if any) with each call to
|
96 |
|
|
load_balance() in this domain when the cpu was busy
|
97 |
|
|
13) # of times in this domain pull_task() was called when busy
|
98 |
|
|
14) # of times in this domain pull_task() was called even though the
|
99 |
|
|
target task was cache-hot when busy
|
100 |
|
|
15) # of times in this domain load_balance() was called but did not
|
101 |
|
|
find a busier queue while the cpu was busy
|
102 |
|
|
16) # of times in this domain a busier queue was found while the cpu
|
103 |
|
|
was busy but no busier group was found
|
104 |
|
|
|
105 |
|
|
17) # of times in this domain load_balance() was called when the
|
106 |
|
|
cpu was just becoming idle
|
107 |
|
|
18) # of times in this domain load_balance() checked but found the
|
108 |
|
|
load did not require balancing when the cpu was just becoming idle
|
109 |
|
|
19) # of times in this domain load_balance() tried to move one or more
|
110 |
|
|
tasks and failed, when the cpu was just becoming idle
|
111 |
|
|
20) sum of imbalances discovered (if any) with each call to
|
112 |
|
|
load_balance() in this domain when the cpu was just becoming idle
|
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
|
142 |
|
|
|
143 |
|
|
/proc//schedstat
|
144 |
|
|
----------------
|
145 |
|
|
schedstats also adds a new /proc/
|
146 |
|
|
the same information on a per-process level. There are three fields in
|
147 |
|
|
this file correlating for that process to:
|
148 |
|
|
1) time spent on the cpu
|
149 |
|
|
2) time spent waiting on a runqueue
|
150 |
|
|
3) # of timeslices run on this cpu
|
151 |
|
|
|
152 |
|
|
A program could be easily written to make use of these extra fields to
|
153 |
|
|
report on how well a particular process or set of processes is faring
|
154 |
|
|
under the scheduler's policies. A simple version of such a program is
|
155 |
|
|
available at
|
156 |
|
|
http://eaglet.rain.com/rick/linux/schedstat/v12/latency.c
|