Repository navigation
Expand file tree
/
Copy pathprofile.8
More file actions
130 lines (122 loc) · 4.44 KB
/
Copy pathprofile.8
File metadata and controls
130 lines (122 loc) · 4.44 KB
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
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
.TH profile 8 "2016-07-17" "USER COMMANDS"
.SH NAME
profile \- Profile CPU usage by sampling stack traces. Uses Linux eBPF/bcc.
.SH SYNOPSIS
.B profile [\-adfh] [\-p PID] [\-U | \-k] [\-F FREQUENCY]
.B [\-\-stack\-storage\-size COUNT] [duration]
.SH DESCRIPTION
This is a CPU profiler. It works by taking samples of stack traces at timed
intervals. It will help you understand and quantify CPU usage: which code is
executing, and by how much, including both user-level and kernel code.
By default this samples at 49 Hertz (samples per second), across all CPUs.
This frequency can be tuned using a command line option. The reason for 49, and
not 50, is to avoid lock-step sampling.
This is also an efficient profiler, as stack traces are frequency counted in
kernel context, rather than passing each stack to user space for frequency
counting there. Only the unique stacks and counts are passed to user space
at the end of the profile, greatly reducing the kernel<->user transfer.
.SH REQUIREMENTS
CONFIG_BPF and bcc.
This also requires Linux 4.9+ (BPF_PROG_TYPE_PERF_EVENT support). See tools/old
for an older version that may work on Linux 4.6 - 4.8.
.SH OPTIONS
.TP
\-h
Print usage message.
.TP
\-p PID
Trace this process ID only (filtered in-kernel). Without this, all CPUs are
profiled.
.TP
\-F frequency
Frequency to sample stacks (default 49).
.TP
\-f
Print output in folded stack format.
.TP
\-d
Include an output delimiter between kernel and user stacks (either "--", or,
in folded mode, "-").
.TP
\-U
Show stacks from user space only (no kernel space stacks).
.TP
\-K
Show stacks from kernel space only (no user space stacks).
.TP
\-\-stack-storage-size COUNT
The maximum number of unique stack traces that the kernel will count (default
10240). If the sampled count exceeds this, a warning will be printed.
.TP
duration
Duration to trace, in seconds.
.SH EXAMPLES
.TP
Profile (sample) stack traces
768A
system-wide at 49 Hertz (samples per second) until Ctrl-C:
#
.B profile
.TP
Profile for 5 seconds only:
#
.B profile 5
.TP
Profile at 99 Hertz for 5 seconds only:
#
.B profile -F 99 5
.TP
Profile PID 181 only:
#
.B profile -p 181
.TP
Profile for 5 seconds and output in folded stack format (suitable as input for flame graphs), including a delimiter between kernel and user stacks:
#
.B profile -df 5
.TP
Profile kernel stacks only:
#
.B profile -K
.SH DEBUGGING
See "[unknown]" frames with bogus addresses? This can happen for different
reasons. Your best approach is to get Linux perf to work first, and then to
try this tool. Eg, "perf record \-F 49 \-a \-g \-\- sleep 1; perf script", and
to check for unknown frames there.
The most common reason for "[unknown]" frames is that the target software has
not been compiled
with frame pointers, and so we can't use that simple method for walking the
stack. The fix in that case is to use software that does have frame pointers,
eg, gcc -fno-omit-frame-pointer, or Java's -XX:+PreserveFramePointer.
Another reason for "[unknown]" frames is JIT compilers, which don't use a
traditional symbol table. The fix in that case is to populate a
/tmp/perf-PID.map file with the symbols, which this tool should read. How you
do this depends on the runtime (Java, Node.js).
If you seem to have unrelated samples in the output, check for other
sampling or tracing tools that may be running. The current version of this
tool can include their events if profiling happened concurrently. Those
samples may be filtered in a future version.
.SH OVERHEAD
This is an efficient profiler, as stack traces are frequency counted in
kernel context, and only the unique stacks and their counts are passed to
user space. Contrast this with the current "perf record -F 99 -a" method
of profiling, which writes each sample to user space (via a ring buffer),
and then to the file system (perf.data), which must be post-processed.
This uses perf_event_open to setup a timer which is instrumented by BPF,
and for efficiency it does not initialize the perf ring buffer, so the
redundant perf samples are not collected.
It's expected that the overhead while sampling at 49 Hertz (the default),
across all CPUs, should be negligible. If you increase the sample rate, the
overhead might begin to be measurable.
.SH SOURCE
This is from bcc.
.IP
https://github.com/iovisor/bcc
.PP
Also look in the bcc distribution for a companion _examples.txt file containing
example usage, output, and commentary for this tool.
.SH OS
Linux
.SH STABILITY
Unstable - in development.
.SH AUTHOR
Brendan Gregg
.SH SEE ALSO
offcputime(8)