| Eric Fiselier | 133a720 | 2017-04-18 07:17:20 +0000 | [diff] [blame] | 1 | # Benchmark Tools |
| 2 | |
| 3 | ## compare_bench.py |
| 4 | |
| 5 | The `compare_bench.py` utility which can be used to compare the result of benchmarks. |
| 6 | The program is invoked like: |
| 7 | |
| 8 | ``` bash |
| 9 | $ compare_bench.py <old-benchmark> <new-benchmark> [benchmark options]... |
| 10 | ``` |
| 11 | |
| 12 | Where `<old-benchmark>` and `<new-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. |
| 13 | |
| Eric Fiselier | 1903976 | 2018-01-18 04:23:01 +0000 | [diff] [blame] | 14 | `[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. |
| 15 | |
| Eric Fiselier | 133a720 | 2017-04-18 07:17:20 +0000 | [diff] [blame] | 16 | The sample output using the JSON test files under `Inputs/` gives: |
| 17 | |
| 18 | ``` bash |
| 19 | $ ./compare_bench.py ./gbench/Inputs/test1_run1.json ./gbench/Inputs/test1_run2.json |
| 20 | Comparing ./gbench/Inputs/test1_run1.json to ./gbench/Inputs/test1_run2.json |
| Eric Fiselier | 1903976 | 2018-01-18 04:23:01 +0000 | [diff] [blame] | 21 | Benchmark Time CPU Time Old Time New CPU Old CPU New |
| 22 | ------------------------------------------------------------------------------------------------------------- |
| 23 | BM_SameTimes +0.0000 +0.0000 10 10 10 10 |
| 24 | BM_2xFaster -0.5000 -0.5000 50 25 50 25 |
| 25 | BM_2xSlower +1.0000 +1.0000 50 100 50 100 |
| 26 | BM_1PercentFaster -0.0100 -0.0100 100 99 100 99 |
| 27 | BM_1PercentSlower +0.0100 +0.0100 100 101 100 101 |
| 28 | BM_10PercentFaster -0.1000 -0.1000 100 90 100 90 |
| 29 | BM_10PercentSlower +0.1000 +0.1000 100 110 100 110 |
| 30 | BM_100xSlower +99.0000 +99.0000 100 10000 100 10000 |
| 31 | BM_100xFaster -0.9900 -0.9900 10000 100 10000 100 |
| 32 | BM_10PercentCPUToTime +0.1000 -0.1000 100 110 100 90 |
| 33 | BM_ThirdFaster -0.3333 -0.3334 100 67 100 67 |
| 34 | BM_BadTimeUnit -0.9000 +0.2000 0 0 0 1 |
| Eric Fiselier | 133a720 | 2017-04-18 07:17:20 +0000 | [diff] [blame] | 35 | ``` |
| 36 | |
| Eric Fiselier | 1903976 | 2018-01-18 04:23:01 +0000 | [diff] [blame] | 37 | As you can note, the values in `Time` and `CPU` columns are calculated as `(new - old) / |old|`. |
| 38 | |
| Eric Fiselier | 133a720 | 2017-04-18 07:17:20 +0000 | [diff] [blame] | 39 | When a benchmark executable is run, the raw output from the benchmark is printed in real time to stdout. The sample output using `benchmark/basic_test` for both arguments looks like: |
| 40 | |
| 41 | ``` |
| 42 | ./compare_bench.py test/basic_test test/basic_test --benchmark_filter=BM_empty.* |
| Eric Fiselier | 1903976 | 2018-01-18 04:23:01 +0000 | [diff] [blame] | 43 | RUNNING: test/basic_test --benchmark_filter=BM_empty.* --benchmark_out=/tmp/tmpN7LF3a |
| 44 | Run on (8 X 4000 MHz CPU s) |
| 45 | 2017-11-07 23:28:36 |
| 46 | --------------------------------------------------------------------- |
| Eric Fiselier | 133a720 | 2017-04-18 07:17:20 +0000 | [diff] [blame] | 47 | Benchmark Time CPU Iterations |
| Eric Fiselier | 1903976 | 2018-01-18 04:23:01 +0000 | [diff] [blame] | 48 | --------------------------------------------------------------------- |
| 49 | BM_empty 4 ns 4 ns 170178757 |
| 50 | BM_empty/threads:8 1 ns 7 ns 103868920 |
| 51 | BM_empty_stop_start 0 ns 0 ns 1000000000 |
| 52 | BM_empty_stop_start/threads:8 0 ns 0 ns 1403031720 |
| 53 | RUNNING: /test/basic_test --benchmark_filter=BM_empty.* --benchmark_out=/tmp/tmplvrIp8 |
| 54 | Run on (8 X 4000 MHz CPU s) |
| 55 | 2017-11-07 23:28:38 |
| 56 | --------------------------------------------------------------------- |
| Eric Fiselier | 133a720 | 2017-04-18 07:17:20 +0000 | [diff] [blame] | 57 | Benchmark Time CPU Iterations |
| Eric Fiselier | 1903976 | 2018-01-18 04:23:01 +0000 | [diff] [blame] | 58 | --------------------------------------------------------------------- |
| 59 | BM_empty 4 ns 4 ns 169534855 |
| 60 | BM_empty/threads:8 1 ns 7 ns 104188776 |
| 61 | BM_empty_stop_start 0 ns 0 ns 1000000000 |
| 62 | BM_empty_stop_start/threads:8 0 ns 0 ns 1404159424 |
| 63 | Comparing ../build/test/basic_test to ../build/test/basic_test |
| 64 | Benchmark Time CPU Time Old Time New CPU Old CPU New |
| 65 | --------------------------------------------------------------------------------------------------------------------- |
| 66 | BM_empty -0.0048 -0.0049 4 4 4 4 |
| 67 | BM_empty/threads:8 -0.0123 -0.0054 1 1 7 7 |
| 68 | BM_empty_stop_start -0.0000 -0.0000 0 0 0 0 |
| 69 | BM_empty_stop_start/threads:8 -0.0029 +0.0001 0 0 0 0 |
| 70 | |
| Eric Fiselier | 133a720 | 2017-04-18 07:17:20 +0000 | [diff] [blame] | 71 | ``` |
| 72 | |
| Eric Fiselier | 1903976 | 2018-01-18 04:23:01 +0000 | [diff] [blame] | 73 | As you can note, the values in `Time` and `CPU` columns are calculated as `(new - old) / |old|`. |
| Eric Fiselier | 133a720 | 2017-04-18 07:17:20 +0000 | [diff] [blame] | 74 | Obviously this example doesn't give any useful output, but it's intended to show the output format when 'compare_bench.py' needs to run benchmarks. |
| Eric Fiselier | 1903976 | 2018-01-18 04:23:01 +0000 | [diff] [blame] | 75 | |
| 76 | ## compare.py |
| 77 | |
| 78 | The `compare.py` can be used to compare the result of benchmarks. |
| 79 | There are three modes of operation: |
| 80 | |
| 81 | 1. Just compare two benchmarks, what `compare_bench.py` did. |
| 82 | The program is invoked like: |
| 83 | |
| 84 | ``` bash |
| 85 | $ compare.py benchmarks <benchmark_baseline> <benchmark_contender> [benchmark options]... |
| 86 | ``` |
| 87 | 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. |
| 88 | |
| 89 | `[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. |
| 90 | |
| 91 | Example output: |
| 92 | ``` |
| 93 | $ ./compare.py benchmarks ./a.out ./a.out |
| 94 | RUNNING: ./a.out --benchmark_out=/tmp/tmprBT5nW |
| 95 | Run on (8 X 4000 MHz CPU s) |
| 96 | 2017-11-07 21:16:44 |
| 97 | ------------------------------------------------------ |
| 98 | Benchmark Time CPU Iterations |
| 99 | ------------------------------------------------------ |
| 100 | BM_memcpy/8 36 ns 36 ns 19101577 211.669MB/s |
| 101 | BM_memcpy/64 76 ns 76 ns 9412571 800.199MB/s |
| 102 | BM_memcpy/512 84 ns 84 ns 8249070 5.64771GB/s |
| 103 | BM_memcpy/1024 116 ns 116 ns 6181763 8.19505GB/s |
| 104 | BM_memcpy/8192 643 ns 643 ns 1062855 11.8636GB/s |
| 105 | BM_copy/8 222 ns 222 ns 3137987 34.3772MB/s |
| 106 | BM_copy/64 1608 ns 1608 ns 432758 37.9501MB/s |
| 107 | BM_copy/512 12589 ns 12589 ns 54806 38.7867MB/s |
| 108 | BM_copy/1024 25169 ns 25169 ns 27713 38.8003MB/s |
| 109 | BM_copy/8192 201165 ns 201112 ns 3486 38.8466MB/s |
| 110 | RUNNING: ./a.out --benchmark_out=/tmp/tmpt1wwG_ |
| 111 | Run on (8 X 4000 MHz CPU s) |
| 112 | 2017-11-07 21:16:53 |
| 113 | ------------------------------------------------------ |
| 114 | Benchmark Time CPU Iterations |
| 115 | ------------------------------------------------------ |
| 116 | BM_memcpy/8 36 ns 36 ns 19397903 211.255MB/s |
| 117 | BM_memcpy/64 73 ns 73 ns 9691174 839.635MB/s |
| 118 | BM_memcpy/512 85 ns 85 ns 8312329 5.60101GB/s |
| 119 | BM_memcpy/1024 118 ns 118 ns 6438774 8.11608GB/s |
| 120 | BM_memcpy/8192 656 ns 656 ns 1068644 11.6277GB/s |
| 121 | BM_copy/8 223 ns 223 ns 3146977 34.2338MB/s |
| 122 | BM_copy/64 1611 ns 1611 ns 435340 37.8751MB/s |
| 123 | BM_copy/512 12622 ns 12622 ns 54818 38.6844MB/s |
| 124 | BM_copy/1024 25257 ns 25239 ns 27779 38.6927MB/s |
| 125 | BM_copy/8192 205013 ns 205010 ns 3479 38.108MB/s |
| 126 | Comparing ./a.out to ./a.out |
| 127 | Benchmark Time CPU Time Old Time New CPU Old CPU New |
| 128 | ------------------------------------------------------------------------------------------------------ |
| 129 | BM_memcpy/8 +0.0020 +0.0020 36 36 36 36 |
| 130 | BM_memcpy/64 -0.0468 -0.0470 76 73 76 73 |
| 131 | BM_memcpy/512 +0.0081 +0.0083 84 85 84 85 |
| 132 | BM_memcpy/1024 +0.0098 +0.0097 116 118 116 118 |
| 133 | BM_memcpy/8192 +0.0200 +0.0203 643 656 643 656 |
| 134 | BM_copy/8 +0.0046 +0.0042 222 223 222 223 |
| 135 | BM_copy/64 +0.0020 +0.0020 1608 1611 1608 1611 |
| 136 | BM_copy/512 +0.0027 +0.0026 12589 12622 12589 12622 |
| 137 | BM_copy/1024 +0.0035 +0.0028 25169 25257 25169 25239 |
| 138 | BM_copy/8192 +0.0191 +0.0194 201165 205013 201112 205010 |
| 139 | ``` |
| 140 | |
| 141 | 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. |
| 142 | As you can note, the values in `Time` and `CPU` columns are calculated as `(new - old) / |old|`. |
| 143 | |
| 144 | 2. Compare two different filters of one benchmark |
| 145 | The program is invoked like: |
| 146 | |
| 147 | ``` bash |
| 148 | $ compare.py filters <benchmark> <filter_baseline> <filter_contender> [benchmark options]... |
| 149 | ``` |
| 150 | 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. |
| 151 | |
| 152 | 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. |
| 153 | |
| 154 | `[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. |
| 155 | |
| 156 | Example output: |
| 157 | ``` |
| 158 | $ ./compare.py filters ./a.out BM_memcpy BM_copy |
| 159 | RUNNING: ./a.out --benchmark_filter=BM_memcpy --benchmark_out=/tmp/tmpBWKk0k |
| 160 | Run on (8 X 4000 MHz CPU s) |
| 161 | 2017-11-07 21:37:28 |
| 162 | ------------------------------------------------------ |
| 163 | Benchmark Time CPU Iterations |
| 164 | ------------------------------------------------------ |
| 165 | BM_memcpy/8 36 ns 36 ns 17891491 211.215MB/s |
| 166 | BM_memcpy/64 74 ns 74 ns 9400999 825.646MB/s |
| 167 | BM_memcpy/512 87 ns 87 ns 8027453 5.46126GB/s |
| 168 | BM_memcpy/1024 111 ns 111 ns 6116853 8.5648GB/s |
| 169 | BM_memcpy/8192 657 ns 656 ns 1064679 11.6247GB/s |
| 170 | RUNNING: ./a.out --benchmark_filter=BM_copy --benchmark_out=/tmp/tmpAvWcOM |
| 171 | Run on (8 X 4000 MHz CPU s) |
| 172 | 2017-11-07 21:37:33 |
| 173 | ---------------------------------------------------- |
| 174 | Benchmark Time CPU Iterations |
| 175 | ---------------------------------------------------- |
| 176 | BM_copy/8 227 ns 227 ns 3038700 33.6264MB/s |
| 177 | BM_copy/64 1640 ns 1640 ns 426893 37.2154MB/s |
| 178 | BM_copy/512 12804 ns 12801 ns 55417 38.1444MB/s |
| 179 | BM_copy/1024 25409 ns 25407 ns 27516 38.4365MB/s |
| 180 | BM_copy/8192 202986 ns 202990 ns 3454 38.4871MB/s |
| 181 | Comparing BM_memcpy to BM_copy (from ./a.out) |
| 182 | Benchmark Time CPU Time Old Time New CPU Old CPU New |
| 183 | -------------------------------------------------------------------------------------------------------------------- |
| 184 | [BM_memcpy vs. BM_copy]/8 +5.2829 +5.2812 36 227 36 227 |
| 185 | [BM_memcpy vs. BM_copy]/64 +21.1719 +21.1856 74 1640 74 1640 |
| 186 | [BM_memcpy vs. BM_copy]/512 +145.6487 +145.6097 87 12804 87 12801 |
| 187 | [BM_memcpy vs. BM_copy]/1024 +227.1860 +227.1776 111 25409 111 25407 |
| 188 | [BM_memcpy vs. BM_copy]/8192 +308.1664 +308.2898 657 202986 656 202990 |
| 189 | ``` |
| 190 | |
| 191 | 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. |
| 192 | As you can note, the values in `Time` and `CPU` columns are calculated as `(new - old) / |old|`. |
| 193 | |
| 194 | 3. Compare filter one from benchmark one to filter two from benchmark two: |
| 195 | The program is invoked like: |
| 196 | |
| 197 | ``` bash |
| 198 | $ compare.py filters <benchmark_baseline> <filter_baseline> <benchmark_contender> <filter_contender> [benchmark options]... |
| 199 | ``` |
| 200 | |
| 201 | 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. |
| 202 | |
| 203 | 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. |
| 204 | |
| 205 | `[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. |
| 206 | |
| 207 | Example output: |
| 208 | ``` |
| 209 | $ ./compare.py benchmarksfiltered ./a.out BM_memcpy ./a.out BM_copy |
| 210 | RUNNING: ./a.out --benchmark_filter=BM_memcpy --benchmark_out=/tmp/tmp_FvbYg |
| 211 | Run on (8 X 4000 MHz CPU s) |
| 212 | 2017-11-07 21:38:27 |
| 213 | ------------------------------------------------------ |
| 214 | Benchmark Time CPU Iterations |
| 215 | ------------------------------------------------------ |
| 216 | BM_memcpy/8 37 ns 37 ns 18953482 204.118MB/s |
| 217 | BM_memcpy/64 74 ns 74 ns 9206578 828.245MB/s |
| 218 | BM_memcpy/512 91 ns 91 ns 8086195 5.25476GB/s |
| 219 | BM_memcpy/1024 120 ns 120 ns 5804513 7.95662GB/s |
| 220 | BM_memcpy/8192 664 ns 664 ns 1028363 11.4948GB/s |
| 221 | RUNNING: ./a.out --benchmark_filter=BM_copy --benchmark_out=/tmp/tmpDfL5iE |
| 222 | Run on (8 X 4000 MHz CPU s) |
| 223 | 2017-11-07 21:38:32 |
| 224 | ---------------------------------------------------- |
| 225 | Benchmark Time CPU Iterations |
| 226 | ---------------------------------------------------- |
| 227 | BM_copy/8 230 ns 230 ns 2985909 33.1161MB/s |
| 228 | BM_copy/64 1654 ns 1653 ns 419408 36.9137MB/s |
| 229 | BM_copy/512 13122 ns 13120 ns 53403 37.2156MB/s |
| 230 | BM_copy/1024 26679 ns 26666 ns 26575 36.6218MB/s |
| 231 | BM_copy/8192 215068 ns 215053 ns 3221 36.3283MB/s |
| 232 | Comparing BM_memcpy (from ./a.out) to BM_copy (from ./a.out) |
| 233 | Benchmark Time CPU Time Old Time New CPU Old CPU New |
| 234 | -------------------------------------------------------------------------------------------------------------------- |
| 235 | [BM_memcpy vs. BM_copy]/8 +5.1649 +5.1637 37 230 37 230 |
| 236 | [BM_memcpy vs. BM_copy]/64 +21.4352 +21.4374 74 1654 74 1653 |
| 237 | [BM_memcpy vs. BM_copy]/512 +143.6022 +143.5865 91 13122 91 13120 |
| 238 | [BM_memcpy vs. BM_copy]/1024 +221.5903 +221.4790 120 26679 120 26666 |
| 239 | [BM_memcpy vs. BM_copy]/8192 +322.9059 +323.0096 664 215068 664 215053 |
| 240 | ``` |
| 241 | 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. |
| 242 | As you can note, the values in `Time` and `CPU` columns are calculated as `(new - old) / |old|`. |