search

lemire / FastPFor

The FastPFOR C++ library: Fast integer compression
Apache License 2.0
883 stars 124 forks source link
compression-schemes simd-compression sorted-lists

readme

The FastPFOR C++ library : Fast integer compression

Ubuntu-CI

What is this?

A research library with integer compression schemes. It is broadly applicable to the compression of arrays of 32-bit integers where most integers are small. The library seeks to exploit SIMD instructions (SSE) whenever possible.

This library can decode at least 4 billions of compressed integers per second on most desktop or laptop processors. That is, it can decompress data at a rate of 15 GB/s. This is significantly faster than generic codecs like gzip, LZO, Snappy or LZ4.

It is used by the zsearch engine as well as in GMAP and GSNAP. DuckDB derived some of their code from this library It has been ported to Java, C# and Go. The Java port is used by ClueWeb Tools.

Apache Lucene version 4.6.x uses a compression format derived from our FastPFOR scheme.

Python bindings

Myths

Myth: SIMD compression requires very large blocks of integers (1024 or more).

Fact: This is not true. Our fastest scheme (SIMDBinaryPacking) works over blocks of 128 integers. Another very fast scheme (Stream VByte) works over blocks of four integers.

Myth: SIMD compression means high speed but less compression.

Fact: This is wrong. Some schemes cannot easily be accelerated with SIMD instructions, but many that do compress very well.

Working with sorted lists of integers

If you are working primarily with sorted lists of integers, then you might want to use differential coding. That is you may want to compress the deltas instead of the integers themselves. The current library (fastpfor) is generic and was not optimized for this purpose. However, we have another library designed to compress sorted integer lists:

https://github.com/lemire/SIMDCompressionAndIntersection

This other library (SIMDCompressionAndIntersection) also comes complete with new SIMD-based intersection algorithms.

There is also a C library for differential coding (fast computation of deltas, and recovery from deltas):

https://github.com/lemire/FastDifferentialCoding

Other recommended libraries

Reference and documentation

For a simple example, please see

example.cpp

in the root directory of this project.

Please see:

This library was used by several papers including the following:

It has also inspired related work such as...

License

This code is licensed under Apache License, Version 2.0 (ASL2.0).

Software Requirements

This code requires a compiler supporting C++11. This was a design decision.

It builds under

The code was tested under Windows, Linux and MacOS.

Hardware Requirements

On an x64 platform, your processor should support SSSE3. This includes almost every Intel or AMD processor sold after 2006. (Note: the key schemes require merely SSE2.) Some specific binaries will only run if your processor supports SSE4.1. They have been purely used for specific tests however.

We also support ARM platforms through SIMDe, by wrapping.

Building with CMake

You need cmake. On most linux distributions, you can simply do the following:

  git clone https://github.com/lemire/FastPFor.git
  cd FastPFor
  mkdir build
  cd build
  cmake ..
  cmake --build .

It may be necessary to set the CXX variable. The project is installable (make install works).

To create project files for Microsoft Visual Studio, it might be useful to target 64-bit Windows (e.g., see http://www.cmake.org/cmake/help/v3.0/generator/Visual%20Studio%2012%202013.html).

Multithreaded context

You should not assume that our objects are thread safe. If you have several threads, each thread should have its own IntegerCODEC objects to ensure that there is no concurrency problems.

Why C++11?

With minor changes, all schemes will compile fine under compilers that do not support C++11. And porting the code to C should not be a challenge.

In any case, we already support 3 major C++ compilers so portability is not a major issue.

What if I prefer Java?

Many schemes cannot be efficiently ported to Java. However some have been. Please see:

https://github.com/lemire/JavaFastPFOR

What if I prefer C#?

See CSharpFastPFOR: A C# integer compression library https://github.com/Genbox/CSharpFastPFOR

What if I prefer Go?

See Encoding: Integer Compression Libraries for Go https://github.com/zhenjl/encoding

Testing

If you used CMake to generate the build files, the check target will run the unit tests. For example , if you generated Unix Makefiles

make check

will do it.

Simple benchmark

make codecs
./codecs --clusterdynamic
./codecs --uniformdynamic

Optional : Snappy

Typing "make allallall" will install some testing binaries that depend on Google Snappy. If you want to build these, you need to install Google snappy. You can do so on a recent ubuntu machine as:

sudo apt-get install libsnappy-dev

Processing data files

Typing "make" will generate an "inmemorybenchmark" executable that can process data files.

You can use it to process arrays on (sorted!) integers on disk using the following 32-bit format: 1 unsigned 32-bit integer indicating array length followed by the corresponding number of 32-bit integer. Repeat.

( It is assumed that the integers are sorted.)

Once you have such a binary file somefilename you can process it with our inmemorybenchmark:

./inmemorybenchmark --minlength 10000 somefilename

The "minlength" flag skips short arrays. (Warning: timings over short arrays are unreliable.)

Testing with the Gov2 and ClueWeb09 data sets

As of April 2014, we recommend getting our archive at

http://lemire.me/data/integercompression2014.html

It is the data was used for the following paper:

Daniel Lemire, Leonid Boytsov, Nathan Kurz, SIMD Compression and the Intersection of Sorted Integers, arXiv: 1401.6399, 2014 http://arxiv.org/abs/1401.6399

I used your code and I get segmentation faults

Our code is thoroughly tested.

One common issue is that people do not provide large enough buffers. Some schemes can have such small compression rates that the compressed data generated will be much larger than the input data.

Is any of this code subject to patents?

I (D. Lemire) did not patent anything.

However, we implemented varint-G8UI which was patented by its authors. DO NOT use varint-G8UI if you want to avoid patents.

The rest of the library should be patent-free.

Funding

This work was supported by NSERC grant number 26143.