xref: /linux/tools/perf/Documentation/cpu-and-latency-overheads.txt (revision 4f9786035f9e519db41375818e1d0b5f20da2f10)
1CPU and latency overheads
2-------------------------
3There are two notions of time: wall-clock time and CPU time.
4For a single-threaded program, or a program running on a single-core machine,
5these notions are the same. However, for a multi-threaded/multi-process program
6running on a multi-core machine, these notions are significantly different.
7Each second of wall-clock time we have number-of-cores seconds of CPU time.
8Perf can measure overhead for both of these times (shown in 'overhead' and
9'latency' columns for CPU and wall-clock time correspondingly).
10
11Optimizing CPU overhead is useful to improve 'throughput', while optimizing
12latency overhead is useful to improve 'latency'. It's important to understand
13which one is useful in a concrete situation at hand. For example, the former
14may be useful to improve max throughput of a CI build server that runs on 100%
15CPU utilization, while the latter may be useful to improve user-perceived
16latency of a single interactive program build.
17These overheads may be significantly different in some cases. For example,
18consider a program that executes function 'foo' for 9 seconds with 1 thread,
19and then executes function 'bar' for 1 second with 128 threads (consumes
20128 seconds of CPU time). The CPU overhead is: 'foo' - 6.6%, 'bar' - 93.4%.
21While the latency overhead is: 'foo' - 90%, 'bar' - 10%. If we try to optimize
22running time of the program looking at the (wrong in this case) CPU overhead,
23we would concentrate on the function 'bar', but it can yield only 10% running
24time improvement at best.
25
26By default, perf shows only CPU overhead. To show latency overhead, use
27'perf record --latency' and 'perf report':
28
29-----------------------------------
30Overhead  Latency  Command
31  93.88%   25.79%  cc1
32   1.90%   39.87%  gzip
33   0.99%   10.16%  dpkg-deb
34   0.57%    1.00%  as
35   0.40%    0.46%  sh
36-----------------------------------
37
38To sort by latency overhead, use 'perf report --latency':
39
40-----------------------------------
41Latency  Overhead  Command
42 39.87%     1.90%  gzip
43 25.79%    93.88%  cc1
44 10.16%     0.99%  dpkg-deb
45  4.17%     0.29%  git
46  2.81%     0.11%  objtool
47-----------------------------------
48
49To get insight into the difference between the overheads, you may check
50parallelization histogram with '--sort=latency,parallelism,comm,symbol --hierarchy'
51flags. It shows fraction of (wall-clock) time the workload utilizes different
52numbers of cores ('Parallelism' column). For example, in the following case
53the workload utilizes only 1 core most of the time, but also has some
54highly-parallel phases, which explains significant difference between
55CPU and wall-clock overheads:
56
57-----------------------------------
58  Latency  Overhead     Parallelism / Command / Symbol
59+  56.98%     2.29%     1
60+  16.94%     1.36%     2
61+   4.00%    20.13%     125
62+   3.66%    18.25%     124
63+   3.48%    17.66%     126
64+   3.26%     0.39%     3
65+   2.61%    12.93%     123
66-----------------------------------
67
68By expanding corresponding lines, you may see what commands/functions run
69at the given parallelism level:
70
71-----------------------------------
72  Latency  Overhead     Parallelism / Command / Symbol
73-  56.98%     2.29%     1
74      32.80%     1.32%     gzip
75       4.46%     0.18%     cc1
76       2.81%     0.11%     objtool
77       2.43%     0.10%     dpkg-source
78       2.22%     0.09%     ld
79       2.10%     0.08%     dpkg-genchanges
80-----------------------------------
81
82To see the normal function-level profile for particular parallelism levels
83(number of threads actively running on CPUs), you may use '--parallelism'
84filter. For example, to see the profile only for low parallelism phases
85of a workload use '--latency --parallelism=1-2' flags.
86