| blob | f2d0c497f3fc7eda95d149f157c205947dd19bdd |
1 # Benchmark Tools
3 ## compare.py
5 The `compare.py` can be used to compare the result of benchmarks.
7 ### Dependencies
8 The utility relies on the [scipy](https://www.scipy.org) package which can be installed using pip:
9 ```bash
10 pip3 install -r requirements.txt
11 ```
13 ### Displaying aggregates only
15 The switch `-a` / `--display_aggregates_only` can be used to control the
16 displayment of the normal iterations vs the aggregates. When passed, it will
17 be passthrough to the benchmark binaries to be run, and will be accounted for
18 in the tool itself; only the aggregates will be displayed, but not normal runs.
19 It only affects the display, the separate runs will still be used to calculate
20 the U test.
22 ### Modes of operation
24 There are three modes of operation:
26 1. Just compare two benchmarks
27 The program is invoked like:
29 ``` bash
30 $ compare.py benchmarks <benchmark_baseline> <benchmark_contender> [benchmark options]...
31 ```
32 Where `<benchmark_baseline>` and `<benchmark_contender>` either specify a benchmark executable file, or a JSON output file. The type of the input file is automatically detected. If a benchmark executable is specified then the benchmark is run to obtain the results. Otherwise the results are simply loaded from the output file.
34 `[benchmark options]` will be passed to the benchmarks invocations. They can be anything that binary accepts, be it either normal `--benchmark_*` parameters, or some custom parameters your binary takes.
36 Example output:
37 ```
38 $ ./compare.py benchmarks ./a.out ./a.out
39 RUNNING: ./a.out --benchmark_out=/tmp/tmprBT5nW
40 Run on (8 X 4000 MHz CPU s)
41 2017-11-07 21:16:44
42 ------------------------------------------------------
43 Benchmark Time CPU Iterations
44 ------------------------------------------------------
45 BM_memcpy/8 36 ns 36 ns 19101577 211.669MB/s
46 BM_memcpy/64 76 ns 76 ns 9412571 800.199MB/s
47 BM_memcpy/512 84 ns 84 ns 8249070 5.64771GB/s
48 BM_memcpy/1024 116 ns 116 ns 6181763 8.19505GB/s
49 BM_memcpy/8192 643 ns 643 ns 1062855 11.8636GB/s
50 BM_copy/8 222 ns 222 ns 3137987 34.3772MB/s
51 BM_copy/64 1608 ns 1608 ns 432758 37.9501MB/s
52 BM_copy/512 12589 ns 12589 ns 54806 38.7867MB/s
53 BM_copy/1024 25169 ns 25169 ns 27713 38.8003MB/s
54 BM_copy/8192 201165 ns 201112 ns 3486 38.8466MB/s
55 RUNNING: ./a.out --benchmark_out=/tmp/tmpt1wwG_
56 Run on (8 X 4000 MHz CPU s)
57 2017-11-07 21:16:53
58 ------------------------------------------------------
59 Benchmark Time CPU Iterations
60 ------------------------------------------------------
61 BM_memcpy/8 36 ns 36 ns 19397903 211.255MB/s
62 BM_memcpy/64 73 ns 73 ns 9691174 839.635MB/s
63 BM_memcpy/512 85 ns 85 ns 8312329 5.60101GB/s
64 BM_memcpy/1024 118 ns 118 ns 6438774 8.11608GB/s
65 BM_memcpy/8192 656 ns 656 ns 1068644 11.6277GB/s
66 BM_copy/8 223 ns 223 ns 3146977 34.2338MB/s
67 BM_copy/64 1611 ns 1611 ns 435340 37.8751MB/s
68 BM_copy/512 12622 ns 12622 ns 54818 38.6844MB/s
69 BM_copy/1024 25257 ns 25239 ns 27779 38.6927MB/s
70 BM_copy/8192 205013 ns 205010 ns 3479 38.108MB/s
71 Comparing ./a.out to ./a.out
72 Benchmark Time CPU Time Old Time New CPU Old CPU New
73 ------------------------------------------------------------------------------------------------------
74 BM_memcpy/8 +0.0020 +0.0020 36 36 36 36
75 BM_memcpy/64 -0.0468 -0.0470 76 73 76 73
76 BM_memcpy/512 +0.0081 +0.0083 84 85 84 85
77 BM_memcpy/1024 +0.0098 +0.0097 116 118 116 118
78 BM_memcpy/8192 +0.0200 +0.0203 643 656 643 656
79 BM_copy/8 +0.0046 +0.0042 222 223 222 223
80 BM_copy/64 +0.0020 +0.0020 1608 1611 1608 1611
81 BM_copy/512 +0.0027 +0.0026 12589 12622 12589 12622
82 BM_copy/1024 +0.0035 +0.0028 25169 25257 25169 25239
83 BM_copy/8192 +0.0191 +0.0194 201165 205013 201112 205010
84 ```
86 What it does is for the every benchmark from the first run it looks for the benchmark with exactly the same name in the second run, and then compares the results. If the names differ, the benchmark is omitted from the diff.
87 As you can note, the values in `Time` and `CPU` columns are calculated as `(new - old) / |old|`.
89 2. Compare two different filters of one benchmark
90 The program is invoked like:
92 ``` bash
93 $ compare.py filters <benchmark> <filter_baseline> <filter_contender> [benchmark options]...
94 ```
95 Where `<benchmark>` either specify a benchmark executable file, or a JSON output file. The type of the input file is automatically detected. If a benchmark executable is specified then the benchmark is run to obtain the results. Otherwise the results are simply loaded from the output file.
97 Where `<filter_baseline>` and `<filter_contender>` are the same regex filters that you would pass to the `[--benchmark_filter=<regex>]` parameter of the benchmark binary.
99 `[benchmark options]` will be passed to the benchmarks invocations. They can be anything that binary accepts, be it either normal `--benchmark_*` parameters, or some custom parameters your binary takes.
101 Example output:
102 ```
103 $ ./compare.py filters ./a.out BM_memcpy BM_copy
104 RUNNING: ./a.out --benchmark_filter=BM_memcpy --benchmark_out=/tmp/tmpBWKk0k
105 Run on (8 X 4000 MHz CPU s)
106 2017-11-07 21:37:28
107 ------------------------------------------------------
108 Benchmark Time CPU Iterations
109 ------------------------------------------------------
110 BM_memcpy/8 36 ns 36 ns 17891491 211.215MB/s
111 BM_memcpy/64 74 ns 74 ns 9400999 825.646MB/s
112 BM_memcpy/512 87 ns 87 ns 8027453 5.46126GB/s
113 BM_memcpy/1024 111 ns 111 ns 6116853 8.5648GB/s
114 BM_memcpy/8192 657 ns 656 ns 1064679 11.6247GB/s
115 RUNNING: ./a.out --benchmark_filter=BM_copy --benchmark_out=/tmp/tmpAvWcOM
116 Run on (8 X 4000 MHz CPU s)
117 2017-11-07 21:37:33
118 ----------------------------------------------------
119 Benchmark Time CPU Iterations
120 ----------------------------------------------------
121 BM_copy/8 227 ns 227 ns 3038700 33.6264MB/s
122 BM_copy/64 1640 ns 1640 ns 426893 37.2154MB/s
123 BM_copy/512 12804 ns 12801 ns 55417 38.1444MB/s
124 BM_copy/1024 25409 ns 25407 ns 27516 38.4365MB/s
125 BM_copy/8192 202986 ns 202990 ns 3454 38.4871MB/s
126 Comparing BM_memcpy to BM_copy (from ./a.out)
127 Benchmark Time CPU Time Old Time New CPU Old CPU New
128 --------------------------------------------------------------------------------------------------------------------
129 [BM_memcpy vs. BM_copy]/8 +5.2829 +5.2812 36 227 36 227
130 [BM_memcpy vs. BM_copy]/64 +21.1719 +21.1856 74 1640 74 1640
131 [BM_memcpy vs. BM_copy]/512 +145.6487 +145.6097 87 12804 87 12801
132 [BM_memcpy vs. BM_copy]/1024 +227.1860 +227.1776 111 25409 111 25407
133 [BM_memcpy vs. BM_copy]/8192 +308.1664 +308.2898 657 202986 656 202990
134 ```
136 As you can see, it applies filter to the benchmarks, both when running the benchmark, and before doing the diff. And to make the diff work, the matches are replaced with some common string. Thus, you can compare two different benchmark families within one benchmark binary.
137 As you can note, the values in `Time` and `CPU` columns are calculated as `(new - old) / |old|`.
139 3. Compare filter one from benchmark one to filter two from benchmark two:
140 The program is invoked like:
142 ``` bash
143 $ compare.py filters <benchmark_baseline> <filter_baseline> <benchmark_contender> <filter_contender> [benchmark options]...
144 ```
146 Where `<benchmark_baseline>` and `<benchmark_contender>` either specify a benchmark executable file, or a JSON output file. The type of the input file is automatically detected. If a benchmark executable is specified then the benchmark is run to obtain the results. Otherwise the results are simply loaded from the output file.
148 Where `<filter_baseline>` and `<filter_contender>` are the same regex filters that you would pass to the `[--benchmark_filter=<regex>]` parameter of the benchmark binary.
150 `[benchmark options]` will be passed to the benchmarks invocations. They can be anything that binary accepts, be it either normal `--benchmark_*` parameters, or some custom parameters your binary takes.
152 Example output:
153 ```
154 $ ./compare.py benchmarksfiltered ./a.out BM_memcpy ./a.out BM_copy
155 RUNNING: ./a.out --benchmark_filter=BM_memcpy --benchmark_out=/tmp/tmp_FvbYg
156 Run on (8 X 4000 MHz CPU s)
157 2017-11-07 21:38:27
158 ------------------------------------------------------
159 Benchmark Time CPU Iterations
160 ------------------------------------------------------
161 BM_memcpy/8 37 ns 37 ns 18953482 204.118MB/s
162 BM_memcpy/64 74 ns 74 ns 9206578 828.245MB/s
163 BM_memcpy/512 91 ns 91 ns 8086195 5.25476GB/s
164 BM_memcpy/1024 120 ns 120 ns 5804513 7.95662GB/s
165 BM_memcpy/8192 664 ns 664 ns 1028363 11.4948GB/s
166 RUNNING: ./a.out --benchmark_filter=BM_copy --benchmark_out=/tmp/tmpDfL5iE
167 Run on (8 X 4000 MHz CPU s)
168 2017-11-07 21:38:32
169 ----------------------------------------------------
170 Benchmark Time CPU Iterations
171 ----------------------------------------------------
172 BM_copy/8 230 ns 230 ns 2985909 33.1161MB/s
173 BM_copy/64 1654 ns 1653 ns 419408 36.9137MB/s
174 BM_copy/512 13122 ns 13120 ns 53403 37.2156MB/s
175 BM_copy/1024 26679 ns 26666 ns 26575 36.6218MB/s
176 BM_copy/8192 215068 ns 215053 ns 3221 36.3283MB/s
177 Comparing BM_memcpy (from ./a.out) to BM_copy (from ./a.out)
178 Benchmark Time CPU Time Old Time New CPU Old CPU New
179 --------------------------------------------------------------------------------------------------------------------
180 [BM_memcpy vs. BM_copy]/8 +5.1649 +5.1637 37 230 37 230
181 [BM_memcpy vs. BM_copy]/64 +21.4352 +21.4374 74 1654 74 1653
182 [BM_memcpy vs. BM_copy]/512 +143.6022 +143.5865 91 13122 91 13120
183 [BM_memcpy vs. BM_copy]/1024 +221.5903 +221.4790 120 26679 120 26666
184 [BM_memcpy vs. BM_copy]/8192 +322.9059 +323.0096 664 215068 664 215053
185 ```
186 This is a mix of the previous two modes, two (potentially different) benchmark binaries are run, and a different filter is applied to each one.
187 As you can note, the values in `Time` and `CPU` columns are calculated as `(new - old) / |old|`.
189 ### U test
191 If there is a sufficient repetition count of the benchmarks, the tool can do
192 a [U Test](https://en.wikipedia.org/wiki/Mann%E2%80%93Whitney_U_test), of the
193 null hypothesis that it is equally likely that a randomly selected value from
194 one sample will be less than or greater than a randomly selected value from a
195 second sample.
197 If the calculated p-value is below this value is lower than the significance
198 level alpha, then the result is said to be statistically significant and the
199 null hypothesis is rejected. Which in other words means that the two benchmarks
200 aren't identical.
202 **WARNING**: requires **LARGE** (no less than 9) number of repetitions to be
203 meaningful!