Filter for improving compression of typed binary data.
Bitshuffle is an algorithm that rearranges typed, binary data for improving compression, as well as a python/C package that implements this algorithm within the Numpy framework.
The library can be used along side HDF5 to compress and decompress datasets and
is integrated through the dynamically loaded filters framework. Bitshuffle
is HDF5 filter number 32008.
Algorithmically, Bitshuffle is closely related to HDF5's Shuffle filter except it operates at the bit level instead of the byte level. Arranging a typed data array in to a matrix with the elements as the rows and the bits within the elements as the columns, Bitshuffle "transposes" the matrix, such that all the least-significant-bits are in a row, etc. This transpose is performed within blocks of data roughly 8kB long [1].
This does not in itself compress data, only rearranges it for more efficient compression. To perform the actual compression you will need a compression library. Bitshuffle has been designed to be well matched to Marc Lehmann's LZF as well as LZ4 and ZSTD. Note that because Bitshuffle modifies the data at the bit level, sophisticated entropy reducing compression libraries such as GZIP and BZIP are unlikely to achieve significantly better compression than simpler and faster duplicate-string-elimination algorithms such as LZF, LZ4 and ZSTD. Bitshuffle thus includes routines (and HDF5 filter options) to apply LZ4 and ZSTD compression to each block after shuffling [2].
The Bitshuffle algorithm relies on neighbouring elements of a dataset being highly correlated to improve data compression. Any correlations that span at least 24 elements of the dataset may be exploited to improve compression.
Bitshuffle was designed with performance in mind. On most machines the time required for Bitshuffle+LZ4 is insignificant compared to the time required to read or write the compressed data to disk. Because it is able to exploit the SSE and AVX instruction sets present on modern Intel and AMD processors, on these machines compression is only marginally slower than an out-of-cache memory copy. On modern x86 processors you can expect Bitshuffle to have a throughput of roughly 1 byte per clock cycle, and on the Haswell generation of Intel processors (2013) and later, you can expect up to 2 bytes per clock cycle. In addition, Bitshuffle is parallelized using OpenMP.
As a bonus, Bitshuffle ships with a dynamically loaded version of
h5py's LZF compression filter, such that the filter can be transparently
used outside of python and in command line utilities such as h5dump.
| [1] | Chosen to fit comfortably within L1 cache as well as be well matched window of the LZF compression library. |
| [2] | Over applying bitshuffle to the full dataset then applying LZ4/ZSTD compression, this has the tremendous advantage that the block is already in the L1 cache. |
Bitshuffle might be right for your application if:
- You need to compress typed binary data.
- Your data is arranged such that adjacent elements over the fastest varying index of your dataset are similar (highly correlated).
- A special case of the previous point is if you are only exercising a subset of the bits in your data-type, as is often true of integer data.
- You need both high compression ratios and high performance.
Comparing Bitshuffle to other compression algorithms and HDF5 filters:
- Bitshuffle is less general than many other compression algorithms. To achieve good compression ratios, consecutive elements of your data must be highly correlated.
- For the right datasets, Bitshuffle is one of the few compression algorithms that promises both high throughput and high compression ratios.
- Bitshuffle should have roughly the same throughput as Shuffle, but may obtain higher compression ratios.
- The MAFISC filter actually includes something similar to Bitshuffle as one of its prefilters, However, MAFICS's emphasis is on obtaining high compression ratios at all costs, sacrificing throughput.
Installation requires python 2.7+ or 3.3+, HDF5 1.8.4 or later, HDF5 for python (h5py), Numpy and Cython. Bitshuffle is linked against HDF5. To use the dynamically loaded HDF5 filter requires HDF5 1.8.11 or later. If ZSTD support is enabled the ZSTD repo needs to pulled into bitshuffle before installation with:
git submodule update --init
To install bitshuffle:
python setup.py install [--h5plugin [--h5plugin-dir=spam] --zstd]
To get finer control of installation options, including whether to compile
with OpenMP multi-threading, copy the setup.cfg.example to setup.cfg
and edit the values therein.
If using the dynamically loaded HDF5 filter (which gives you access to the
Bitshuffle and LZF filters outside of python), set the environment variable
HDF5_PLUGIN_PATH to the value of --h5plugin-dir or use HDF5's default
search location of /usr/local/hdf5/lib/plugin.
ZSTD support is enabled with --zstd.
If you get an error about missing source files when building the extensions, try upgrading setuptools. There is a weird bug where setuptools prior to 0.7 doesn't work properly with Cython in some cases.
The bitshuffle module contains routines for shuffling and unshuffling Numpy arrays.
If installed with the dynamically loaded filter plugins, Bitshuffle can be used
in conjunction with HDF5 both inside and outside of python, in the same way as
any other filter; simply by specifying the filter number 32008. Otherwise
the filter will be available only within python and only after importing
bitshuffle.h5. Reading Bitshuffle encoded datasets will be transparent.
The filter can be added to new datasets either through the h5py low level
interface or through the convenience functions provided in
bitshuffle.h5. See the docstrings and unit tests for examples. For h5py
version 2.5.0 and later Bitshuffle can be added to new datasets through the
high level interface, as in the example below.
The compression algorithm can be configured using the filter_opts in bitshuffle.h5.create_dataset(). LZ4 is chosen with: (BLOCK_SIZE, h5.H5_COMPRESS_LZ4) and ZSTD with: (BLOCK_SIZE, h5.H5_COMPRESS_ZSTD, COMP_LVL). See test_h5filter.py for an example.
import h5py
import numpy
import bitshuffle.h5
print(h5py.__version__) # >= '2.5.0'
f = h5py.File(filename, "w")
# block_size = 0 let Bitshuffle choose its value
block_size = 0
dataset = f.create_dataset(
"data",
(100, 100, 100),
compression=bitshuffle.h5.H5FILTER,
compression_opts=(block_size, bitshuffle.h5.H5_COMPRESS_LZ4),
dtype='float32',
)
# create some random data
array = numpy.random.rand(100, 100, 100)
array = array.astype('float32')
dataset[:] = array
f.close()
If you wish to use Bitshuffle in your C program and would prefer not to use the
HDF5 dynamically loaded filter, the C library in the src/ directory is
self-contained and complete.
You can use Bitshuffle even in Java and the routines for shuffling and unshuffling are ported into snappy-java. To use the routines, you need to add the following dependency to your pom.xml:
<dependency> <groupId>org.xerial.snappy</groupId> <artifactId>snappy-java</artifactId> <version>1.1.3-M1</version> </dependency>
First, import org.xerial.snapy.BitShuffle in your Java code:
import org.xerial.snappy.BitShuffle;
Then, you use them like this:
int[] data = new int[] {1, 3, 34, 43, 34};
byte[] shuffledData = BitShuffle.bitShuffle(data);
int[] result = BitShuffle.bitUnShuffleIntArray(shuffledData);
The conda package can be build via:
conda build conda-recipe
Here are a few tips to help you get the most out of Bitshuffle:
- For multi-dimensional datasets, order your data such that the fastest varying dimension is the one over which your data is most correlated (have values that change the least), or fake this using chunks.
- To achieve the highest throughput, use a data type that is 64 bytes or smaller. If you have a very large compound data type, consider adding a dimension to your datasets instead.
- To make full use of the SSE2 instruction set, use a data type whose size is a multiple of 2 bytes. For the AVX2 instruction set, use a data type whose size is a multiple of 4 bytes.
Bitshuffle was initially described in http://dx.doi.org/10.1016/j.ascom.2015.07.002, pre-print available at http://arxiv.org/abs/1503.00638.