1# Benchmark Tools 2 3## compare.py 4 5The `compare.py` can be used to compare the result of benchmarks. 6 7**NOTE**: the utility relies on the scipy package which can be installed using [these instructions](https://www.scipy.org/install.html). 8 9### Displaying aggregates only 10 11The switch `-a` / `--display_aggregates_only` can be used to control the 12displayment of the normal iterations vs the aggregates. When passed, it will 13be passthrough to the benchmark binaries to be run, and will be accounted for 14in the tool itself; only the aggregates will be displayed, but not normal runs. 15It only affects the display, the separate runs will still be used to calculate 16the U test. 17 18### Modes of operation 19 20There are three modes of operation: 21 221. Just compare two benchmarks 23The program is invoked like: 24 25``` bash 26$ compare.py benchmarks <benchmark_baseline> <benchmark_contender> [benchmark options]... 27``` 28Where `<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. 29 30`[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. 31 32Example output: 33``` 34$ ./compare.py benchmarks ./a.out ./a.out 35RUNNING: ./a.out --benchmark_out=/tmp/tmprBT5nW 36Run on (8 X 4000 MHz CPU s) 372017-11-07 21:16:44 38------------------------------------------------------ 39Benchmark Time CPU Iterations 40------------------------------------------------------ 41BM_memcpy/8 36 ns 36 ns 19101577 211.669MB/s 42BM_memcpy/64 76 ns 76 ns 9412571 800.199MB/s 43BM_memcpy/512 84 ns 84 ns 8249070 5.64771GB/s 44BM_memcpy/1024 116 ns 116 ns 6181763 8.19505GB/s 45BM_memcpy/8192 643 ns 643 ns 1062855 11.8636GB/s 46BM_copy/8 222 ns 222 ns 3137987 34.3772MB/s 47BM_copy/64 1608 ns 1608 ns 432758 37.9501MB/s 48BM_copy/512 12589 ns 12589 ns 54806 38.7867MB/s 49BM_copy/1024 25169 ns 25169 ns 27713 38.8003MB/s 50BM_copy/8192 201165 ns 201112 ns 3486 38.8466MB/s 51RUNNING: ./a.out --benchmark_out=/tmp/tmpt1wwG_ 52Run on (8 X 4000 MHz CPU s) 532017-11-07 21:16:53 54------------------------------------------------------ 55Benchmark Time CPU Iterations 56------------------------------------------------------ 57BM_memcpy/8 36 ns 36 ns 19397903 211.255MB/s 58BM_memcpy/64 73 ns 73 ns 9691174 839.635MB/s 59BM_memcpy/512 85 ns 85 ns 8312329 5.60101GB/s 60BM_memcpy/1024 118 ns 118 ns 6438774 8.11608GB/s 61BM_memcpy/8192 656 ns 656 ns 1068644 11.6277GB/s 62BM_copy/8 223 ns 223 ns 3146977 34.2338MB/s 63BM_copy/64 1611 ns 1611 ns 435340 37.8751MB/s 64BM_copy/512 12622 ns 12622 ns 54818 38.6844MB/s 65BM_copy/1024 25257 ns 25239 ns 27779 38.6927MB/s 66BM_copy/8192 205013 ns 205010 ns 3479 38.108MB/s 67Comparing ./a.out to ./a.out 68Benchmark Time CPU Time Old Time New CPU Old CPU New 69------------------------------------------------------------------------------------------------------ 70BM_memcpy/8 +0.0020 +0.0020 36 36 36 36 71BM_memcpy/64 -0.0468 -0.0470 76 73 76 73 72BM_memcpy/512 +0.0081 +0.0083 84 85 84 85 73BM_memcpy/1024 +0.0098 +0.0097 116 118 116 118 74BM_memcpy/8192 +0.0200 +0.0203 643 656 643 656 75BM_copy/8 +0.0046 +0.0042 222 223 222 223 76BM_copy/64 +0.0020 +0.0020 1608 1611 1608 1611 77BM_copy/512 +0.0027 +0.0026 12589 12622 12589 12622 78BM_copy/1024 +0.0035 +0.0028 25169 25257 25169 25239 79BM_copy/8192 +0.0191 +0.0194 201165 205013 201112 205010 80``` 81 82What 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. 83As you can note, the values in `Time` and `CPU` columns are calculated as `(new - old) / |old|`. 84 852. Compare two different filters of one benchmark 86The program is invoked like: 87 88``` bash 89$ compare.py filters <benchmark> <filter_baseline> <filter_contender> [benchmark options]... 90``` 91Where `<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. 92 93Where `<filter_baseline>` and `<filter_contender>` are the same regex filters that you would pass to the `[--benchmark_filter=<regex>]` parameter of the benchmark binary. 94 95`[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. 96 97Example output: 98``` 99$ ./compare.py filters ./a.out BM_memcpy BM_copy 100RUNNING: ./a.out --benchmark_filter=BM_memcpy --benchmark_out=/tmp/tmpBWKk0k 101Run on (8 X 4000 MHz CPU s) 1022017-11-07 21:37:28 103------------------------------------------------------ 104Benchmark Time CPU Iterations 105------------------------------------------------------ 106BM_memcpy/8 36 ns 36 ns 17891491 211.215MB/s 107BM_memcpy/64 74 ns 74 ns 9400999 825.646MB/s 108BM_memcpy/512 87 ns 87 ns 8027453 5.46126GB/s 109BM_memcpy/1024 111 ns 111 ns 6116853 8.5648GB/s 110BM_memcpy/8192 657 ns 656 ns 1064679 11.6247GB/s 111RUNNING: ./a.out --benchmark_filter=BM_copy --benchmark_out=/tmp/tmpAvWcOM 112Run on (8 X 4000 MHz CPU s) 1132017-11-07 21:37:33 114---------------------------------------------------- 115Benchmark Time CPU Iterations 116---------------------------------------------------- 117BM_copy/8 227 ns 227 ns 3038700 33.6264MB/s 118BM_copy/64 1640 ns 1640 ns 426893 37.2154MB/s 119BM_copy/512 12804 ns 12801 ns 55417 38.1444MB/s 120BM_copy/1024 25409 ns 25407 ns 27516 38.4365MB/s 121BM_copy/8192 202986 ns 202990 ns 3454 38.4871MB/s 122Comparing BM_memcpy to BM_copy (from ./a.out) 123Benchmark Time CPU Time Old Time New CPU Old CPU New 124-------------------------------------------------------------------------------------------------------------------- 125[BM_memcpy vs. BM_copy]/8 +5.2829 +5.2812 36 227 36 227 126[BM_memcpy vs. BM_copy]/64 +21.1719 +21.1856 74 1640 74 1640 127[BM_memcpy vs. BM_copy]/512 +145.6487 +145.6097 87 12804 87 12801 128[BM_memcpy vs. BM_copy]/1024 +227.1860 +227.1776 111 25409 111 25407 129[BM_memcpy vs. BM_copy]/8192 +308.1664 +308.2898 657 202986 656 202990 130``` 131 132As 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. 133As you can note, the values in `Time` and `CPU` columns are calculated as `(new - old) / |old|`. 134 1353. Compare filter one from benchmark one to filter two from benchmark two: 136The program is invoked like: 137 138``` bash 139$ compare.py filters <benchmark_baseline> <filter_baseline> <benchmark_contender> <filter_contender> [benchmark options]... 140``` 141 142Where `<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. 143 144Where `<filter_baseline>` and `<filter_contender>` are the same regex filters that you would pass to the `[--benchmark_filter=<regex>]` parameter of the benchmark binary. 145 146`[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. 147 148Example output: 149``` 150$ ./compare.py benchmarksfiltered ./a.out BM_memcpy ./a.out BM_copy 151RUNNING: ./a.out --benchmark_filter=BM_memcpy --benchmark_out=/tmp/tmp_FvbYg 152Run on (8 X 4000 MHz CPU s) 1532017-11-07 21:38:27 154------------------------------------------------------ 155Benchmark Time CPU Iterations 156------------------------------------------------------ 157BM_memcpy/8 37 ns 37 ns 18953482 204.118MB/s 158BM_memcpy/64 74 ns 74 ns 9206578 828.245MB/s 159BM_memcpy/512 91 ns 91 ns 8086195 5.25476GB/s 160BM_memcpy/1024 120 ns 120 ns 5804513 7.95662GB/s 161BM_memcpy/8192 664 ns 664 ns 1028363 11.4948GB/s 162RUNNING: ./a.out --benchmark_filter=BM_copy --benchmark_out=/tmp/tmpDfL5iE 163Run on (8 X 4000 MHz CPU s) 1642017-11-07 21:38:32 165---------------------------------------------------- 166Benchmark Time CPU Iterations 167---------------------------------------------------- 168BM_copy/8 230 ns 230 ns 2985909 33.1161MB/s 169BM_copy/64 1654 ns 1653 ns 419408 36.9137MB/s 170BM_copy/512 13122 ns 13120 ns 53403 37.2156MB/s 171BM_copy/1024 26679 ns 26666 ns 26575 36.6218MB/s 172BM_copy/8192 215068 ns 215053 ns 3221 36.3283MB/s 173Comparing BM_memcpy (from ./a.out) to BM_copy (from ./a.out) 174Benchmark Time CPU Time Old Time New CPU Old CPU New 175-------------------------------------------------------------------------------------------------------------------- 176[BM_memcpy vs. BM_copy]/8 +5.1649 +5.1637 37 230 37 230 177[BM_memcpy vs. BM_copy]/64 +21.4352 +21.4374 74 1654 74 1653 178[BM_memcpy vs. BM_copy]/512 +143.6022 +143.5865 91 13122 91 13120 179[BM_memcpy vs. BM_copy]/1024 +221.5903 +221.4790 120 26679 120 26666 180[BM_memcpy vs. BM_copy]/8192 +322.9059 +323.0096 664 215068 664 215053 181``` 182This is a mix of the previous two modes, two (potentially different) benchmark binaries are run, and a different filter is applied to each one. 183As you can note, the values in `Time` and `CPU` columns are calculated as `(new - old) / |old|`. 184 185### U test 186 187If there is a sufficient repetition count of the benchmarks, the tool can do 188a [U Test](https://en.wikipedia.org/wiki/Mann%E2%80%93Whitney_U_test), of the 189null hypothesis that it is equally likely that a randomly selected value from 190one sample will be less than or greater than a randomly selected value from a 191second sample. 192 193If the calculated p-value is below this value is lower than the significance 194level alpha, then the result is said to be statistically significant and the 195null hypothesis is rejected. Which in other words means that the two benchmarks 196aren't identical. 197 198**WARNING**: requires **LARGE** (no less than 9) number of repetitions to be 199meaningful! 200