Alexander Gromnitsky's Blog

Comparing Compression

Latest update:

Do you benchmark compression tools (like xz or zstd) on your own data, or do you rely on common wisdom? The best result for an uncompressed 300MB XFS image from the previous post was achieved by bzip2, which is rarely used nowadays. How does one quickly check a chunk of data against N popular compressors?

E.g., an unpacked tarball of Emacs 29.2 source code consists of 6791 files with a total size of 276MB. If you were to distribute it as a .tar.something archive, which compression tool would be the optimal choice? We can easily write a small utility that answers this question.

$ ./comprtest ~/opt/src/emacs/emacs-29.2 | tee table
tar: Removing leading `/' from member names
szip             0.59   56.98        126593557
gzip             9.21   72.70         80335332
compress         3.57   57.45        125217137
bzip2           17.28   78.08         64509672
rzip            17.61   79.50         60336377
lzip           113.61   81.67         53935898
lzop             0.67   57.14        126121462
xz             111.03   81.89         53295220
brotli          13.10   78.14         64336399
zstd             1.13   73.77         77179446

comprtest is a 29 LOC long shell script. The 2nd column here indicates time in seconds, the 3rd column displays 100(1-compressedorig) , representing space saving in % (higher % is better), & the 4th column shows the final result in bytes.

Then we can sort the table by the 3rd column & draw a bar chart:

$ sort -nk3 table | cpp -P | gnuplot -persist

If you're wondering why all of a sudden the C preprocessor becomes part of it, read on.

comprtest expects either a file as an argument or a directory (in which case it creates a plain .tar of it first). Additional optional arguments specify which compressors to use:

$ ./comprtest /usr/libexec/gdb gzip brotli
gzip             0.60   61.17          6054706
brotli           1.17   65.84          5325408

The gist of the script involves looping over a list of compressors:

archivers='szip gzip compress bzip2 rzip lzip lzop xz brotli zstd'
for c in ${@:-$archivers}; do
    echo $c
    case $c in
        szip   ) args='< "$input" > $output' ;;
        rzip   ) args='-k -o $output "$input"' ;;
        brotli ) args='-6 -c "$input" > $output' ;;
        *      ) args='-c "$input" > $output'

    eval "time -p $c $args" 2>&1 | awk '/real/ {print $2}'
    osize=`wc -c < $output`

    echo $isize $osize | awk '{print 100*(1-$2/($1==0?$2:$1))}'
    echo $osize
    rm $output
done | xargs -n4 printf "%-8s  %11.2f  %6.2f  %15d\n"
  • Not every archive tool has gzip-compatible CLI.
  • We are using a default compression level for each tool with the exception of brotli, as its default level 11 is excruciatingly slow.
  • szip is an interface to the Snappy algorithm. Your distro probably doesn't have it in its repos, hence run cargo install szip. Everything else should be available via dnf/apt.

Bar charts are generated by a gnuplot script:

$ cat
$data <<E
#include "/dev/stdin"
set key tmargin
set xtics rotate by -30 left
set y2tics
set ylabel "Seconds"
set y2label "%"
set style data histograms
set style fill solid
plot $data using 2 axis x1y1 title "Time", \
     "" using 3:xticlabels(1) axis x1y2 title "Space saving"

Here is where the C preprocessor comes in handy: without an injected "datablock" it won't be possible to draw a graph with 2 ordinates when reading data from stdin.

In an attempt to demonstrate that xz is not always the best choice, I benchmarked a bunch of XML files (314MB):

$ ./comprtest ~/Downloads/
szip             0.59   63.70        119429565
gzip             7.18   77.59         73724710
compress         4.03   67.17        108015563
bzip2           21.37   83.36         54751478
rzip            17.42   85.93         46304199
lzip           119.70   85.06         49151518
lzop             0.67   63.63        119667058
xz             125.80   85.55         47559464
brotli          13.56   82.52         57509978
zstd             1.07   79.40         67766890

Tags: ойті
Authors: ag