These simple Python scripts benchmark different zlib compression libraries. The graph below shows the performance of methods to compress files into the popular gzip format. Options include Mark Adler's popular zlib, the zlib-ng, the CloudFlare zlib, the classic gzip and libdeflate. Note that libdeflate has outstanding performance, but does not work with streams, so it fills a different niche than the zlib libraries. The simplicity of gzip and its early introduction mean it remains extremely popular. However, for context the graph also shows the performance of the modern zstd compression format, which leverages hardware and concepts that were not available when gz was developed.
To run this script, you would run the following commands:
a_compile.py
b_speed_test.py
The script a_compile.py downloads the silesia corpus and compiles the versions of zlib to be tested. For Linux uses, it will compile zlib-ng and CloudFlare using both the gcc and clang compilers. This assumes you have both compilers installed and they are named gcc and clang.
If you compile on an Apple Silicon CPU (M1, M2), you may want to make sure the compiler created arm64 executables not x86_64 executables that must be run in translation. In my experience, if your run the a_compile.py using a x86_64 version of Python you will get x86_64 executables so you should use a native version of Python. To compile the zlib-dougallj for arm64 you must have a hardware and compilers that supports the udot instruction (e.g. clang 13 or later).
$file ./libdeflate
./libdeflate: Mach-O 64-bit executable arm64
The script b_speed_test.py allows us to compare speed of compression and decompression. Here is the output for the default Silesia corpus.
CompressMethod Level Min Mean Max mb/s %
zstd 1 624 628 633 340 34.69
zstd 2 719 729 735 295 32.86
zstd 3 944 951 957 225 31.47
zstd 4 1020 1028 1033 208 30.93
zstd 5 1735 1750 1759 122 30.22
zstd 6 2116 2127 2146 100 29.70
zstd 7 2824 2841 2872 75 28.97
zstd 8 3448 3501 3532 61 28.69
zstd 9 4397 4442 4481 48 28.42
zstd 10 5387 5446 5498 39 28.11
zstd 11 7233 7250 7267 29 27.97
zstd 12 9495 9751 9906 22 27.75
zstd 13 13751 13842 13930 15 27.43
zstd 14 17115 17228 17329 12 27.19
zstd 15 22535 22687 22926 9 27.01
zstd 16 29581 29904 30080 7 26.29
zstd 17 37803 38170 38403 6 25.78
zstd 18 47500 47928 48446 4 25.34
zstd 19 61459 61742 62495 3 25.13
gzip 1 2511 2539 2569 84 36.50
gzip 2 2679 2711 2736 79 35.44
gzip 3 3310 3332 3347 64 34.46
gzip 4 3431 3469 3506 62 33.52
gzip 5 4465 4508 4537 47 32.62
gzip 6 6365 6420 6476 33 32.19
gzip 7 7610 7704 7773 28 32.05
gzip 8 11146 11209 11344 19 31.93
gzip 9 13990 14169 14345 15 31.91
libdeflate 1 1231 1242 1249 172 34.59
libdeflate 2 1320 1331 1335 161 33.77
libdeflate 3 1449 1468 1481 146 33.34
libdeflate 4 1640 1657 1675 129 32.94
libdeflate 5 1804 1818 1828 117 32.32
libdeflate 6 2174 2207 2231 97 32.05
libdeflate 7 2799 2834 2856 76 31.90
libdeflate 8 8955 8997 9048 24 31.55
libdeflate 9 12273 12390 12463 17 31.01
zlibCFclang 1 1723 1735 1763 123 35.80
zlibCFclang 2 1814 1819 1826 117 35.00
zlibCFclang 3 2047 2064 2074 104 34.37
zlibCFclang 4 2211 2250 2299 96 33.46
zlibCFclang 5 2815 2857 2900 75 32.61
zlibCFclang 6 3595 3626 3651 59 32.35
zlibCFclang 7 4155 4185 4216 51 32.25
zlibCFclang 8 6045 6101 6147 35 32.17
zlibCFclang 9 7987 8085 8168 27 32.15
zlibCFgcc 1 1576 1597 1609 134 35.80
zlibCFgcc 2 1661 1676 1692 128 35.00
zlibCFgcc 3 1896 1907 1918 112 34.37
zlibCFgcc 4 2084 2132 2175 102 33.46
zlibCFgcc 5 2626 2654 2669 81 32.61
zlibCFgcc 6 3366 3431 3467 63 32.35
zlibCFgcc 7 3997 4020 4039 53 32.25
zlibCFgcc 8 5847 5881 5922 36 32.17
zlibCFgcc 9 7842 7919 7964 27 32.15
zlibMadler 1 2584 2607 2641 82 36.45
zlibMadler 2 2779 2794 2816 76 35.39
zlibMadler 3 3359 3389 3419 63 34.43
zlibMadler 4 3485 3518 3560 61 33.50
zlibMadler 5 4665 4702 4717 45 32.63
zlibMadler 6 6603 6680 6724 32 32.19
zlibMadler 7 7953 8053 8138 27 32.05
zlibMadler 8 11823 11893 11991 18 31.94
zlibMadler 9 14943 15091 15301 14 31.92
zlibNGclang 1 972 989 998 218 47.69
zlibNGclang 2 1622 1627 1632 131 35.54
zlibNGclang 3 1954 1961 1968 108 34.22
zlibNGclang 4 2191 2232 2262 97 32.95
zlibNGclang 5 2425 2446 2476 87 32.67
zlibNGclang 6 2919 2948 2970 73 32.51
zlibNGclang 7 4365 4396 4434 49 32.25
zlibNGclang 8 6614 6698 6762 32 32.17
zlibNGclang 9 9017 9132 9214 24 32.15
zlibNGgcc 1 1078 1099 1112 197 47.69
zlibNGgcc 2 1720 1731 1741 123 35.54
zlibNGgcc 3 2043 2061 2067 104 34.22
zlibNGgcc 4 2385 2410 2429 89 32.95
zlibNGgcc 5 2610 2637 2673 81 32.67
zlibNGgcc 6 3101 3117 3150 68 32.51
zlibNGgcc 7 4409 4452 4555 48 32.25
zlibNGgcc 8 6595 6667 6742 32 32.17
zlibNGgcc 9 9139 9239 9365 23 32.15
DecompressMethod Min Mean Max mb/s
zstd 7643 7889 7994 526.86
gzip 70063 70527 70756 190.57
libdeflate 32336 32477 32630 412.92
zlibCFclang 48796 49065 49470 273.63
zlibCFgcc 49240 49328 49369 271.17
zlibMadler 56574 56746 56918 236.01
zlibNGclang 47293 47366 47423 282.33
zlibNGgcc 48661 48738 48804 274.39
You can also test compression of files in any folder you want. Some compression methods are tuned for certain types of data (e.g. English text) and may be relatively worse than other methods for other types of data (e.g. medical images). For example, you can download and test the Canterbury corpus or the Calgary corpus. To test the compression of all the files in a folder, provide the folder name as an argument from the command line:
b_speed_test.py ./MyFolder
For decompression testing, all gz tools are decompressing the same data (addressing a concern by Sebastian Pop that different gzip compressors create different file sizes, and smaller files might be more complicated and therefore slower to extract). In contrast, since zstd is creating files in a different format, it is only decompressing its own files.
Be warned that this script will use a huge amount of disk space. The Silesia coprpus is large, and for the decomrpression test each compression tool contributes a compressed version of this corpus at each compression level (1..9).
- zlib-bench is a perl script.
- pigz-bench-python tests the parallel compressor pigz.
- deflatebench is a Python script for testing gz compression. It attempts to control low level CPU features (like turbo modes that can boost performance for short bursts) that can lead to variability in benchmark timing.