Home / server / Questions / 872749
Accepted
guettli
Asked: 2017-09-09 10:39:04 +0800 CST

bzip2 too slow. Multiple cores are avaible

  • 772

I am running this command:

pg_dumpall | bzip2 > cluster-$(date --iso).sql.bz2

It takes too long. I look at the processes with top. The bzip2 process takes about 95% and postgres 5% of one core. The wa entry is low. This means the disk is not the bottleneck.

What can I do to increase the performance?

Maybe let bzip2 use more cores. The servers has 16 cores.

Or use an alternative to bzip2?

What can I do to increase the performance?

performance

5 Answers

  1. Best Answer

    There are many compression algorithms around, and bzip2 is one of the slower ones. Plain gzip tends to be significantly faster, at usually not much worse compression. When speed is the most important, lzop is my favourite. Poor compression, but oh so fast.

    I decided to have some fun and compare a few algorithms, including their parallel implementations. The input file is the output of pg_dumpall command on my workstation, a 1913 MB SQL file. The hardware is an older quad-core i5. The times are wall-clock times of just the compression. Parallel implementations are set to use all 4 cores. Table sorted by compression speed.

    Algorithm     Compressed size        Compression          Decompression

lzop           398MB    20.8%      4.2s    455.6MB/s     3.1s    617.3MB/s
lz4            416MB    21.7%      4.5s    424.2MB/s     1.6s   1181.3MB/s
brotli (q0)    307MB    16.1%      7.3s    262.1MB/s     4.9s    390.5MB/s
brotli (q1)    234MB    12.2%      8.7s    220.0MB/s     4.9s    390.5MB/s
zstd           266MB    13.9%     11.9s    161.1MB/s     3.5s    539.5MB/s
pigz (x4)      232MB    12.1%     13.1s    146.1MB/s     4.2s    455.6MB/s
gzip           232MB    12.1%     39.1s     48.9MB/s     9.2s    208.0MB/s
lbzip2 (x4)    188MB     9.9%     42.0s     45.6MB/s    13.2s    144.9MB/s
pbzip2 (x4)    189MB     9.9%    117.5s     16.3MB/s    20.1s     95.2MB/s
bzip2          189MB     9.9%    273.4s      7.0MB/s    42.8s     44.7MB/s
pixz (x4)      132MB     6.9%    456.3s      4.2MB/s     7.9s    242.2MB/s
xz             132MB     6.9%   1027.8s      1.9MB/s    17.3s    110.6MB/s
brotli (q11)   141MB     7.4%   4979.2s      0.4MB/s     3.6s    531.6MB/s

    If the 16 cores of your server are idle enough that all can be used for compression, pbzip2 will probably give you a very significant speed-up. But you need more speed still and you can tolerate ~20% larger files, gzip is probably your best bet.

    Update: I added brotli (see TOOGAMs answer) results to the table. brotlis compression quality setting has a very large impact on compression ratio and speed, so I added three settings (q0, q1, and q11). The default is q11, but it is extremely slow, and still worse than xz. q1 looks very good though; the same compression ratio as gzip, but 4-5 times as fast!

    Update: Added lbzip2 (see gmathts comment) and zstd (Johnny's comment) to the table, and sorted it by compression speed. lbzip2 puts the bzip2 family back in the running by compressing three times as fast as pbzip2 with a great compression ratio! zstd also looks reasonable but is beat by brotli (q1) in both ratio and speed.

    My original conclusion that plain gzip is the best bet is starting to look almost silly. Although for ubiquity, it still can't be beat ;)

    • 65

  2. Use pbzip2.

    The manual says:

    pbzip2 is a parallel implementation of the bzip2 block-sorting file compressor that uses pthreads and achieves near-linear speedup on SMP machines. The output of this version is fully compatible with bzip2 v1.0.2 or newer (ie: anything compressed with pbzip2 can be decompressed with bzip2).

    It auto-detects the number of processors you have and creates threads accordingly.

    • 42

  3. You didn't mention an operating system. If Windows, 7-Zip with ZStandard (Releases) is a version of 7-Zip that has been modified to provide support for using all of these algorithms.

    • 8

  4. Use zstd. If it's good enough for Facebook, it's probably good enough for you as well.

    On a more serious note, it's actually pretty good. I use it for everything now because it just works, and it lets you trade speed for ratio on a large scale (most often, speed matters more than size anyway since storage is cheap, but speed is a bottleneck).
    At compression levels that achieve comparable overall compression as bzip2, it's significantly faster, and if you are willing to pay some extra in CPU time, you can almost achieve results similar to LZMA (although then it will be slower than bzip2). At sligthly worse compression ratios, it is much, much faster than bzip2 or any other mainstream alternative.

    Now, your are compressing a SQL dump, which is just about as embarrassingly trivial to compress as it can be. Even the poorest compressors score well on that kind of data.
    So you can run zstd with a lower compression level, which will run dozens of times faster and still achieve 95-99% the same compression on that data.

    As a bonus, if you will be doing this often and want to invest some extra time, you can "train" the zstd compressor ahead of time, which increases both compression ratio and speed. Note that for training to work well, you will need to feed it individual records, not the whole thing. The way the tool works, it expects many small and somewhat similar samples for training, not one huge blob.

    • 2

  5. It looks like adjusting (lowering) the block size can have a significant impact on the compression time.

    Here are some results of the experiment I did on my machine. I used the time command to measure the execution time. input.txt is a ~250mb text file containing arbitrary json records.

    Using the default (biggest) block size (--best merely selects the default behaviour):

    # time cat input.txt | bzip2 --best > input-compressed-best.txt.bz

real    0m48.918s
user    0m48.397s
sys     0m0.767s

    Using the smallest block size (--fast argument):

    # time cat input.txt | bzip2 --fast > input-compressed-fast.txt.bz

real    0m33.859s
user    0m33.571s
sys     0m0.741s

    This was a bit surprising discovery, considering that the documntation says:

    Compression and decompression speed are virtually unaffected by block size

    • 1