[forward] Quick Python zlib vs bz2 benchmark

// http://log.bthomson.com/2011/01/quick-python-gzip-vs-bz2-benchmark.html

Quick Python zlib vs bz2 benchmark

I use the zlib mod­ule a lot on Google App En­gine; often the tiny CPU time for de­com­pres­sion is a good trade­off to save disk space. I was cu­ri­ous how bz2 com­pares so I ran this short bench­mark.

The test file was this plain­text book, a highly-com­press­ible source. Columns are: level, time, bytes un­com­pressed, bytes com­pressed, ratio.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47

% ./bench.zsh    zlib compress 0   6.98ms 640599 640700 1.000 1  21.22ms 640599 274195 2.336 2  25.08ms 640599 261638 2.448 3  34.24ms 640599 249649 2.566 4  36.41ms 640599 241500 2.653 5  54.24ms 640599 232545 2.755 6  77.22ms 640599 228621 2.802 7  87.94ms 640599 228032 2.809 8 112.49ms 640599 227622 2.814 9 113.03ms 640599 227622 2.814    zlib decompress 0   1.54ms 1   6.39ms 2   6.13ms 3   6.02ms 4   6.22ms 5   5.96ms 6   5.94ms 7   5.90ms 8   5.89ms 9   5.94ms    bz2 compress 1 105.30ms 640599 196752 3.256 2 103.42ms 640599 186082 3.443 3 105.40ms 640599 180905 3.541 4 104.95ms 640599 177642 3.606 5 113.12ms 640599 176232 3.635 6 110.45ms 640599 173153 3.700 7 113.06ms 640599 169634 3.776 8 110.27ms 640599 169634 3.776 9 111.43ms 640599 169634 3.776    bz2 decompress 1  36.40ms 2  35.79ms 3  36.35ms 4  36.81ms 5  41.18ms 6  44.86ms 7  48.96ms 8  48.45ms 9  47.95ms

Con­clu­sion: prob­a­bly not worth it. bz2 at level=4 takes about 7 times longer to de­com­press than gzip at level=9 for only a mod­est im­prove­ment in the com­pres­sion ratio from 2.8 to 3.6.

In­ter­est­ingly for write-heavy work­loads bz2 may ac­tu­ally be the bet­ter choice since com­pres­sion time is not much worse than gzip at level=9.

I think it's bet­ter not to use the timeit mod­ule for this kind of bench­mark since in typ­i­cal usage you will just be com­press­ing/de­com­press­ing some given data once. If the op­er­a­tions speed up in re­peat runs due to caching (and they do), that doesn't re­flect typ­i­cal usage. Start­ing a new python process for each test seems to re­duce cache ef­fects.

Any­way, here is the code.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24

import zlib import bz2 import time import sys    level = int(sys.argv[1]) mod = zlib if int(sys.argv[2]) else bz2 is_decompress = int(sys.argv[3])    with open("pg4238.txt") as f:   data = f.read()    if is_decompress:   c_data = mod.compress(data, level)    t = time.time() if is_decompress:   data = mod.decompress(c_data) else:   c_data = mod.compress(data, level)    print level, "%6.02fms" % (1000*(time.time() - t)), if not is_decompress:   print len(data), len(c_data), "%.03f" % (float(len(data))/len(c_data))

?

1 2 3 4 5 6 7 8 9

#!/usr/bin/zsh echo 'zlib compress' for level in {0..9}; do python bench.py $level 1 0; done echo '\nzlib decompress' for level in {0..9}; do python bench.py $level 1 1; done echo '\nbz2 compress' for level in {1..9}; do python bench.py $level 0 0; done echo '\nbz2 decompress' for level in {1..9}; do python bench.py $level 0 1; done