| Author: | Francesc Alted i Abad |
|---|---|
| Contact: | [email protected] |
| URL: | http://blosc.pytables.org |
Blosc [1] 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 is the first compressor (that I'm aware of) that is meant not only to reduce the size of large datasets on-disk or in-memory, but also to accelerate memory-bound computations.
It uses the blocking technique (as described in [2]) to reduce activity on the memory bus as much as possible. In short, this technique works by dividing datasets in blocks that are small enough to fit in caches of modern processors and perform compression / decompression there. It also leverages, if available, SIMD instructions (SSE2) and multi-threading capabilities of CPUs, in order to accelerate the compression / decompression process to a maximum.
You can see some recent benchmarks about Blosc performance in [3]
Blosc is distributed using the MIT license, see LICENSES/BLOSC.txt for details.
| [1] | http://blosc.pytables.org |
| [2] | http://www.pytables.org/docs/CISE-12-2-ScientificPro.pdf |
| [3] | http://blosc.pytables.org/trac/wiki/SyntheticBenchmarks |
Blosc is not like other compressors: it should rather be called a meta-compressor. This is so because it can use different compressors and pre-conditioners (programs that generally improve compression ratio). At any rate, it can also be called a compressor because it happens that it already integrates one compressor and one pre-conditioner, so it can actually work like so.
Currently it uses BloscLZ, a compressor heavily based on FastLZ (http://fastlz.org/), and a highly optimized (it can use SSE2 instructions, if available) Shuffle pre-conditioner. However, different compressors or pre-conditioners may be added in the future.
Blosc is in charge of coordinating the compressor and pre-conditioners so that they can leverage the blocking technique (described above) as well as multi-threaded execution (if several cores are available) automatically. That makes that every compressor and pre-conditioner will work at very high speeds, even if it was not initially designed for doing blocking or multi-threading.
Other advantages of Blosc are:
- Meant for binary data: can take advantage of the type size meta-information for improved compression ratio (using the integrated shuffle pre-conditioner).
- Small overhead on non-compressible data: only a maximum of 16 additional bytes over the source buffer length are needed to compress every input.
- Maximum destination length: contrarily to many other compressors, both compression and decompression routines have support for maximum size lengths for the destination buffer.
- Replacement for memcpy(): it supports a 0 compression level that does not compress at all and only adds 16 bytes of overhead. In this mode Blosc can copy memory usually faster than a plain memcpy().
When taken together, all these features set Blosc apart from other similar solutions.
Blosc consists of the next files (in blosc/ directory):
blosc.h and blosc.c -- the main routines blosclz.h and blosclz.c -- the actual compressor shuffle.h and shuffle.c -- the shuffle code
Just add these files to your project in order to use Blosc. For information on compression and decompression routines, see blosc.h.
To compile using GCC (4.4 or higher recommended) on Unix:
gcc -O3 -msse2 -o myprog myprog.c blosc/*.c -lpthread
Using Windows and MINGW:
gcc -O3 -msse2 -o myprog myprog.c blosc\*.c
Using Windows and MSVC (2008 or higher recommended):
cl /Ox /Femyprog.exe myprog.c blosc\*.c
A simple usage example is the benchmark in the bench/bench.c file. Also, another example for using Blosc as a generic HDF5 filter is in the hdf5/ directory.
I have not tried to compile this with compilers other than GCC, MINGW, Intel ICC or MSVC yet. Please report your experiences with your own platforms.
Go to the test/ directory and issue:
$ make test
These tests are very basic, and only valid for platforms where GNU make/gcc tools are available. If you really want to test Blosc the hard way, look at:
http://blosc.pytables.org/trac/wiki/SyntheticBenchmarks
where instructions on how to intensively test (and benchmark) Blosc are given. If while running these tests you get some error, please report it back!
Blosc can also be built, tested and installed using CMake. The following procedure describes the "out of source" build.
Create the build directory and move into it:
$ mkdir build $ cd build
Configure Blosc in release mode (enable optimizations) specifying the installation directory:
$ cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=INSTALL_DIR \
PATH_TO_BLOSC_SOURCE_DIR
Please note that configuration can also be performed using UI tools provided by CMake (ccmake or cmake-gui):
$ cmake-gui PATH_TO_BLOSC_SOURCE_DIR
Build, test and install Blosc:
$ make $ make test $ make install
The static and dynamic version of the Bloasc library, together with header files, will be installed into the specified INSTALL_DIR.
Blosc has an official wrapper for Python. See:
https://github.com/FrancescAlted/python-blosc
For those that want to use Blosc as a filter in the HDF5 library, there is a sample implementation in the hdf5/ directory.
There is an official mailing list for Blosc at:
[email protected] http://groups.google.es/group/blosc
I'd like to thank the PyTables community that have collaborated in the exhaustive testing of Blosc. With an aggregate amount of more than 300 TB of different datasets compressed and decompressed successfully, I can say that Blosc is pretty safe now and ready for production purposes. Also, Valentin Haenel did a terrific work fixing typos and improving docs and the plotting script.
Enjoy data!