JavaFastPFOR: A simple integer compression library in Java
This code is released under the Apache License Version 2.0 http://www.apache.org/licenses/.
What does this do?
It is a library to compress and uncompress arrays of integers very fast. The assumption is that most (but not all) values in your array use much less than 32 bits, or that the gaps between the integers use much less than 32 bits. These sort of arrays often come up when using differential coding in databases and information retrieval (e.g., in inverted indexes or column stores).
Please note that random integers are not compressible, by this library or by any other means. If you ever had the means of systematically compressing random integers, you could compress any data source to nothing, by recursive application of your technique.
This library can decompress integers at a rate of over 1.2 billions per second (4.5 GB/s). It is significantly faster than generic codecs (such as Snappy, LZ4 and so on) when compressing arrays of integers.
Part of this library has been integrated in Parquet (http://parquet.io/). A modified version of the library is included in the search engine Terrier (http://terrier.org/). This libary is used by ClueWeb Tools (https://github.com/lintool/clueweb).
This library inspired a compression scheme used by Apache Lucene (e.g., see http://lucene.apache.org/core/4_6_1/core/org/apache/lucene/util/PForDeltaDocIdSet.html ).
It is a java port of the fastpfor C++ library (https://github.com/lemire/FastPFor). There is also a Go port (https://github.com/reducedb/encoding). The C++ library is used by the zsearch engine (http://victorparmar.github.com/zsearch/) as well as in GMAP and GSNAP (http://research-pub.gene.com/gmap/).
Really simple usage:
IntegratedIntCompressor iic = new IntegratedIntCompressor(); int data = ... ; // to be compressed int compressed = iic.compress(data); // compressed array int recov = iic.uncompress(compressed); // equals to data
For more examples, see example.java.
Some CODECs ("integrated codecs") assume that the integers are in sorted orders and use differential coding (they compress deltas). They can be found in the package me.lemire.integercompression.differential. Most others do not.
Maven central repository
Using this code in your own project is easy with maven, just add the following code in your pom.xml file:
<dependencies> <dependency> <groupId>me.lemire.integercompression</groupId> <artifactId>JavaFastPFOR</artifactId> <version>[0.1,)</version> </dependency> </dependencies>
Naturally, you should replace "version" by the version you desire.
You can also download JavaFastPFOR from the Maven central repository: http://repo1.maven.org/maven2/me/lemire/integercompression/JavaFastPFOR/
We found no library that implemented state-of-the-art integer coding techniques such as Binary Packing, NewPFD, OptPFD, Variable Byte, Simple 9 and so on in Java. We wrote one.
Some codecs are thread-safe while others are not. For this reason, it is best to use one codec per thread. The memory usage of a codec instance is small in any case.
Nevertheless, if you want to reuse codec instances, note that by convention, unless the documentation of a codec specify that it is not thread-safe, then it can be assumed to be thread-safe.
with contributions by
- the Terrier team (Matteo Catena, Craig Macdonald, Saúl Vargas and Iadh Ounis)
- Di Wu, http://www.facebook.com/diwu1989
- Stefan Ackermann, https://github.com/Stivo
- Samit Roy, https://github.com/roysamit
How does it compare to the Kamikaze PForDelta library?
In our tests, Kamikaze PForDelta is slower than our implementations. See the benchmarkresults directory for some results.
A recent Java compiler. Java 7 or better is recommended.
Good instructions on installing Java 7 on Linux:
How fast is it?
Compile the code and execute me.lemire.integercompression.benchmarktools.Benchmark.
I recommend running all the benchmarks with the "-server" flag on a desktop machine.
Speed is always reported in millions of integers per second.
For Maven users
For ant users
If you use Apache ant, please try this:
$ ant Benchmark
$ ant Benchmark -Dbenchmark.target=BenchmarkBitPacking
Want to read more?
This library was a key ingredient in the best paper at ECIR 2014 :
Matteo Catena, Craig Macdonald, Iadh Ounis, On Inverted Index Compression for Search Engine Efficiency, Lecture Notes in Computer Science 8416 (ECIR 2014), 2014. http://dx.doi.org/10.1007/978-3-319-06028-6_30
We wrote several research papers documenting many of the CODECs implemented here:
- Daniel Lemire, Leonid Boytsov, Nathan Kurz, SIMD Compression and the Intersection of Sorted Integers, Software Practice & Experience (to appear) http://arxiv.org/abs/1401.6399
- Daniel Lemire and Leonid Boytsov, Decoding billions of integers per second through vectorization, Software Practice & Experience 45 (1), 2015. http://arxiv.org/abs/1209.2137 http://onlinelibrary.wiley.com/doi/10.1002/spe.2203/abstract
- Jeff Plaisance, Nathan Kurz, Daniel Lemire, Vectorized VByte Decoding, International Symposium on Web Algorithms 2015, 2015. http://arxiv.org/abs/1503.07387
- Wayne Xin Zhao, Xudong Zhang, Daniel Lemire, Dongdong Shan, Jian-Yun Nie, Hongfei Yan, Ji-Rong Wen, A General SIMD-based Approach to Accelerating Compression Algorithms, ACM Transactions on Information Systems 33 (3), 2015. http://arxiv.org/abs/1502.01916
Ikhtear Sharif wrote his M.Sc. thesis on this library:
Ikhtear Sharif, Performance Evaluation of Fast Integer Compression Techniques Over Tables, M.Sc. thesis, UNB 2013. http://lemire.me/fr/documents/thesis/IkhtearThesis.pdf
He also posted his slides online: http://www.slideshare.net/ikhtearSharif/ikhtear-defense
Other recommended libraries
TurboPFor is a C library that offers lots of interesting optimizations and Java wrappers. Well worth checking!
This work was supported by NSERC grant number 26143.