Author: | Francesc Alted |
---|---|
Author: | Valentin Haenel |
Contact: | [email protected] |
Contact: | [email protected] |
Github: | https://github.com/Blosc/python-blosc |
URL: | http://python-blosc.blosc.org |
Travis CI: | |
Appveyor: | |
PyPi: | |
Anaconda: | |
Gitter: |
Blosc (http://blosc.org) is a high performance compressor optimized for binary data. It has been designed to transmit data to the processor cache faster than the traditional, non-compressed, direct memory fetch approach via a memcpy() OS call.
Blosc works well for compressing numerical arrays that contains data with relatively low entropy, like sparse data, time series, grids with regular-spaced values, etc.
python-blosc a Python package that wraps Blosc. python-blosc supports Python 3.6 or higher versions.
The Blosc community has adopted a Code of Conduct that we expect project participants to adhere to. Please read the full text so that you can understand what actions will and will not be tolerated.
You can install binary packages with conda
:
$ conda install -c conda-forge python-blosc
Or, install it as a typical Python source package (requires c-compiler and
Python headers) from PyPi using pip
:
$ pip install blosc
The Sphinx based documentation is here:
Also, some examples are available on python-blosc wiki page:
http://github.com/blosc/python-blosc/wiki
Lastly, here is the recording and the slides from the talk "Compress me stupid" at the EuroPython 2014.
If you need more control, there are different ways to compile python-blosc, depending if you want to link with an already installed Blosc library or not.
python-blosc comes with the Blosc sources with it and can be built with:
$ python -m pip install -r requirements-dev.txt
$ python setup.py build_ext --inplace
Any codec can be enabled (=1) or disabled (=0) on this build-path with the appropriate OS environment variables INCLUDE_LZ4, INCLUDE_SNAPPY, INCLUDE_ZLIB, and INCLUDE_ZLIB. By default all the codecs in Blosc are enabled except Snappy (due to some issues with C++ with the gcc toolchain).
Compiler specific optimisations are automatically enabled by inspecting the CPU flags building Blosc. They can be manually disabled by setting the following environmental variables: DISABLE_BLOSC_SSE2 and DISABLE_BLOSC_AVX2.
setuptools is limited to using the compiler specified in the environment variable CC which on posix systems is usually gcc. This often causes trouble with the Snappy codec, which is written in C++, and as a result Snappy is no longer compiled by default. This problem is not known to affect MSVC or clang. Snappy is considered optional in Blosc as its compression performance is below that of the other codecs.
That's all. You can proceed with testing section now.
This approach uses pre-built, fully optimized versions of Blosc built via CMake.
Go to https://github.com/Blosc/c-blosc/releases and download and install the C-Blosc library. Then, you can tell python-blosc where is the C-Blosc library in a couple of ways:
Using an environment variable:
$ BLOSC_DIR=/usr/local (or "set BLOSC_DIR=\blosc" on Win)
$ export BLOSC_DIR (not needed on Win)
$ python setup.py build_clib
$ python setup.py build_ext --inplace
Using a flag:
$ python setup.py build_clib
$ python setup.py build_ext --inplace --blosc=/usr/local
After compiling, you can quickly check that the package is sane by
running the doctests in blosc/test.py
:
$ PYTHONPATH=. (or "set PYTHONPATH=." on Win)
$ export PYTHONPATH=. (not needed on Win)
$ python blosc/test.py (add -v for verbose mode)
Or alternatively, you can use the third-party nosetests
script:
$ nosetests --with-doctest (add -v for verbose mode)
Once installed, you can re-run the tests at any time with:
$ python -c "import blosc; blosc.test()"
If curious, you may want to run a small benchmark that compares a plain NumPy array copy against compression through different compressors in your Blosc build:
$ PYTHONPATH=. python bench/compress_ptr.py
Just to whet your appetite, here are the results for an Intel Xeon E5-2695 v3 @ 2.30GHz, running Python 3.5, CentOS 7, but YMMV (and will vary!):
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= python-blosc version: 1.5.1.dev0 Blosc version: 1.11.2 ($Date:: 2017-01-27 #$) Compressors available: ['blosclz', 'lz4', 'lz4hc', 'snappy', 'zlib', 'zstd'] Compressor library versions: BloscLZ: 1.0.5 LZ4: 1.7.5 Snappy: 1.1.1 Zlib: 1.2.7 Zstd: 1.1.2 Python version: 3.5.2 |Continuum Analytics, Inc.| (default, Jul 2 2016, 17:53:06) [GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] Platform: Linux-3.10.0-327.18.2.el7.x86_64-x86_64 (#1 SMP Thu May 12 11:03:55 UTC 2016) Linux dist: CentOS Linux 7.2.1511 Processor: x86_64 Byte-ordering: little Detected cores: 56 Number of threads to use by default: 4 -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= Creating NumPy arrays with 10**8 int64/float64 elements: *** ctypes.memmove() *** Time for memcpy(): 0.276 s (2.70 GB/s) Times for compressing/decompressing with clevel=5 and 24 threads *** the arange linear distribution *** *** blosclz , noshuffle *** 0.382 s (1.95 GB/s) / 0.300 s (2.48 GB/s) Compr. ratio: 1.0x *** blosclz , shuffle *** 0.042 s (17.77 GB/s) / 0.027 s (27.18 GB/s) Compr. ratio: 57.1x *** blosclz , bitshuffle *** 0.094 s (7.94 GB/s) / 0.041 s (18.28 GB/s) Compr. ratio: 74.0x *** lz4 , noshuffle *** 0.156 s (4.79 GB/s) / 0.052 s (14.30 GB/s) Compr. ratio: 2.0x *** lz4 , shuffle *** 0.033 s (22.58 GB/s) / 0.034 s (22.03 GB/s) Compr. ratio: 68.6x *** lz4 , bitshuffle *** 0.059 s (12.63 GB/s) / 0.053 s (14.18 GB/s) Compr. ratio: 33.1x *** lz4hc , noshuffle *** 0.443 s (1.68 GB/s) / 0.070 s (10.62 GB/s) Compr. ratio: 2.0x *** lz4hc , shuffle *** 0.102 s (7.31 GB/s) / 0.029 s (25.42 GB/s) Compr. ratio: 97.5x *** lz4hc , bitshuffle *** 0.206 s (3.62 GB/s) / 0.038 s (19.85 GB/s) Compr. ratio: 180.5x *** snappy , noshuffle *** 0.154 s (4.84 GB/s) / 0.056 s (13.28 GB/s) Compr. ratio: 2.0x *** snappy , shuffle *** 0.044 s (16.89 GB/s) / 0.047 s (15.95 GB/s) Compr. ratio: 17.4x *** snappy , bitshuffle *** 0.064 s (11.58 GB/s) / 0.061 s (12.26 GB/s) Compr. ratio: 18.2x *** zlib , noshuffle *** 1.172 s (0.64 GB/s) / 0.135 s (5.50 GB/s) Compr. ratio: 5.3x *** zlib , shuffle *** 0.260 s (2.86 GB/s) / 0.086 s (8.67 GB/s) Compr. ratio: 120.8x *** zlib , bitshuffle *** 0.262 s (2.84 GB/s) / 0.094 s (7.96 GB/s) Compr. ratio: 260.1x *** zstd , noshuffle *** 0.973 s (0.77 GB/s) / 0.093 s (8.00 GB/s) Compr. ratio: 7.8x *** zstd , shuffle *** 0.093 s (7.97 GB/s) / 0.023 s (32.71 GB/s) Compr. ratio: 156.7x *** zstd , bitshuffle *** 0.115 s (6.46 GB/s) / 0.029 s (25.60 GB/s) Compr. ratio: 320.6x *** the linspace linear distribution *** *** blosclz , noshuffle *** 0.341 s (2.19 GB/s) / 0.291 s (2.56 GB/s) Compr. ratio: 1.0x *** blosclz , shuffle *** 0.132 s (5.65 GB/s) / 0.023 s (33.10 GB/s) Compr. ratio: 2.0x *** blosclz , bitshuffle *** 0.166 s (4.50 GB/s) / 0.036 s (20.89 GB/s) Compr. ratio: 2.8x *** lz4 , noshuffle *** 0.142 s (5.26 GB/s) / 0.028 s (27.07 GB/s) Compr. ratio: 1.0x *** lz4 , shuffle *** 0.093 s (8.01 GB/s) / 0.030 s (24.87 GB/s) Compr. ratio: 3.4x *** lz4 , bitshuffle *** 0.102 s (7.31 GB/s) / 0.039 s (19.13 GB/s) Compr. ratio: 5.3x *** lz4hc , noshuffle *** 0.700 s (1.06 GB/s) / 0.044 s (16.77 GB/s) Compr. ratio: 1.1x *** lz4hc , shuffle *** 0.203 s (3.67 GB/s) / 0.021 s (36.22 GB/s) Compr. ratio: 8.6x *** lz4hc , bitshuffle *** 0.342 s (2.18 GB/s) / 0.028 s (26.50 GB/s) Compr. ratio: 14.2x *** snappy , noshuffle *** 0.271 s (2.75 GB/s) / 0.274 s (2.72 GB/s) Compr. ratio: 1.0x *** snappy , shuffle *** 0.099 s (7.54 GB/s) / 0.042 s (17.55 GB/s) Compr. ratio: 4.2x *** snappy , bitshuffle *** 0.127 s (5.86 GB/s) / 0.043 s (17.20 GB/s) Compr. ratio: 6.1x *** zlib , noshuffle *** 1.525 s (0.49 GB/s) / 0.158 s (4.70 GB/s) Compr. ratio: 1.6x *** zlib , shuffle *** 0.346 s (2.15 GB/s) / 0.098 s (7.59 GB/s) Compr. ratio: 10.7x *** zlib , bitshuffle *** 0.420 s (1.78 GB/s) / 0.104 s (7.20 GB/s) Compr. ratio: 18.0x *** zstd , noshuffle *** 1.061 s (0.70 GB/s) / 0.096 s (7.79 GB/s) Compr. ratio: 1.9x *** zstd , shuffle *** 0.203 s (3.68 GB/s) / 0.052 s (14.21 GB/s) Compr. ratio: 14.2x *** zstd , bitshuffle *** 0.251 s (2.97 GB/s) / 0.047 s (15.84 GB/s) Compr. ratio: 22.2x *** the random distribution *** *** blosclz , noshuffle *** 0.340 s (2.19 GB/s) / 0.285 s (2.61 GB/s) Compr. ratio: 1.0x *** blosclz , shuffle *** 0.091 s (8.21 GB/s) / 0.017 s (44.29 GB/s) Compr. ratio: 3.9x *** blosclz , bitshuffle *** 0.080 s (9.27 GB/s) / 0.029 s (26.12 GB/s) Compr. ratio: 6.1x *** lz4 , noshuffle *** 0.150 s (4.95 GB/s) / 0.027 s (28.05 GB/s) Compr. ratio: 2.4x *** lz4 , shuffle *** 0.068 s (11.02 GB/s) / 0.029 s (26.03 GB/s) Compr. ratio: 4.5x *** lz4 , bitshuffle *** 0.063 s (11.87 GB/s) / 0.054 s (13.70 GB/s) Compr. ratio: 6.2x *** lz4hc , noshuffle *** 0.645 s (1.15 GB/s) / 0.019 s (39.22 GB/s) Compr. ratio: 3.5x *** lz4hc , shuffle *** 0.257 s (2.90 GB/s) / 0.022 s (34.62 GB/s) Compr. ratio: 5.1x *** lz4hc , bitshuffle *** 0.128 s (5.80 GB/s) / 0.029 s (25.52 GB/s) Compr. ratio: 6.2x *** snappy , noshuffle *** 0.164 s (4.54 GB/s) / 0.048 s (15.46 GB/s) Compr. ratio: 2.2x *** snappy , shuffle *** 0.082 s (9.09 GB/s) / 0.043 s (17.39 GB/s) Compr. ratio: 4.3x *** snappy , bitshuffle *** 0.071 s (10.48 GB/s) / 0.046 s (16.08 GB/s) Compr. ratio: 5.0x *** zlib , noshuffle *** 1.223 s (0.61 GB/s) / 0.093 s (7.97 GB/s) Compr. ratio: 4.0x *** zlib , shuffle *** 0.636 s (1.17 GB/s) / 0.126 s (5.89 GB/s) Compr. ratio: 5.5x *** zlib , bitshuffle *** 0.327 s (2.28 GB/s) / 0.109 s (6.81 GB/s) Compr. ratio: 6.2x *** zstd , noshuffle *** 1.432 s (0.52 GB/s) / 0.103 s (7.27 GB/s) Compr. ratio: 4.2x *** zstd , shuffle *** 0.388 s (1.92 GB/s) / 0.031 s (23.71 GB/s) Compr. ratio: 5.9x *** zstd , bitshuffle *** 0.127 s (5.86 GB/s) / 0.033 s (22.77 GB/s) Compr. ratio: 6.4x
Also, Blosc works quite well on ARM processors (even without NEON support yet):
-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= python-blosc version: 1.4.4 Blosc version: 1.11.2 ($Date:: 2017-01-27 #$) Compressors available: ['blosclz', 'lz4', 'lz4hc', 'snappy', 'zlib', 'zstd'] Compressor library versions: BloscLZ: 1.0.5 LZ4: 1.7.5 Snappy: 1.1.1 Zlib: 1.2.8 Zstd: 1.1.2 Python version: 3.6.0 (default, Dec 31 2016, 21:20:16) [GCC 4.9.2] Platform: Linux-3.4.113-sun8i-armv7l (#50 SMP PREEMPT Mon Nov 14 08:41:55 CET 2016) Linux dist: debian 9.0 Processor: not recognized Byte-ordering: little Detected cores: 4 Number of threads to use by default: 4 -=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-= *** ctypes.memmove() *** Time for memcpy(): 0.015 s (93.57 MB/s) Times for compressing/decompressing with clevel=5 and 4 threads *** user input *** *** blosclz , noshuffle *** 0.015 s (89.93 MB/s) / 0.010 s (138.32 MB/s) Compr. ratio: 2.7x *** blosclz , shuffle *** 0.023 s (60.25 MB/s) / 0.012 s (112.71 MB/s) Compr. ratio: 2.3x *** blosclz , bitshuffle *** 0.018 s (77.63 MB/s) / 0.021 s (66.76 MB/s) Compr. ratio: 7.3x *** lz4 , noshuffle *** 0.008 s (177.14 MB/s) / 0.009 s (159.00 MB/s) Compr. ratio: 3.6x *** lz4 , shuffle *** 0.010 s (131.29 MB/s) / 0.012 s (117.69 MB/s) Compr. ratio: 3.5x *** lz4 , bitshuffle *** 0.015 s (89.97 MB/s) / 0.022 s (63.62 MB/s) Compr. ratio: 8.4x *** lz4hc , noshuffle *** 0.071 s (19.30 MB/s) / 0.007 s (186.64 MB/s) Compr. ratio: 8.6x *** lz4hc , shuffle *** 0.079 s (17.30 MB/s) / 0.014 s (95.99 MB/s) Compr. ratio: 6.2x *** lz4hc , bitshuffle *** 0.062 s (22.23 MB/s) / 0.027 s (51.53 MB/s) Compr. ratio: 9.7x *** snappy , noshuffle *** 0.008 s (173.87 MB/s) / 0.009 s (148.77 MB/s) Compr. ratio: 4.4x *** snappy , shuffle *** 0.011 s (123.22 MB/s) / 0.016 s (85.16 MB/s) Compr. ratio: 4.4x *** snappy , bitshuffle *** 0.015 s (89.02 MB/s) / 0.021 s (64.87 MB/s) Compr. ratio: 6.2x *** zlib , noshuffle *** 0.047 s (29.26 MB/s) / 0.011 s (121.83 MB/s) Compr. ratio: 14.7x *** zlib , shuffle *** 0.080 s (17.20 MB/s) / 0.022 s (63.61 MB/s) Compr. ratio: 9.4x *** zlib , bitshuffle *** 0.059 s (23.50 MB/s) / 0.033 s (41.10 MB/s) Compr. ratio: 10.5x *** zstd , noshuffle *** 0.113 s (12.21 MB/s) / 0.011 s (124.64 MB/s) Compr. ratio: 15.6x *** zstd , shuffle *** 0.154 s (8.92 MB/s) / 0.026 s (52.56 MB/s) Compr. ratio: 9.9x *** zstd , bitshuffle *** 0.116 s (11.86 MB/s) / 0.036 s (38.40 MB/s) Compr. ratio: 11.4x
For details on the ARM benchmark see: Blosc#105
In case you find your own results interesting, please report them back to the authors!
The software is licenses under a 3-Clause BSD licsense. A copy of the python-blosc license can be found in LICENSE.txt. A copy of all licenses can be found in LICENSES/.
Discussion about this module is welcome in the Blosc list:
http://groups.google.es/group/blosc
Enjoy data!