As a reference, several fast compression algorithms were tested and compared on a Core i7-3930K CPU @ 4.5GHz, using [lzbench], an open-source in-memory benchmark by @inikep compiled with GCC 5.4.0, with the [Silesia compression corpus].
Speed vs Compression trade-off is configurable by small increments. Decompression speed is preserved and remains roughly the same at all settings, a property shared by most LZ compression algorithms, such as [zlib] or lzma.
The following tests were run on a Core i7-3930K CPU @ 4.5GHz, using [lzbench], an open-source in-memory benchmark by @inikep compiled with GCC 5.2.1, on the [Silesia compression corpus].
Previous charts provide results applicable to typical file and stream scenarios (several MB). Small data comes with different perspectives. The smaller the amount of data to compress, the more difficult it is to achieve any significant compression.
This problem is common to many compression algorithms. The reason is, compression algorithms learn from past data how to compress future data. But at the beginning of a new file, there is no "past" to build upon.
To solve this situation, Zstd offers a __training mode__, which can be used to tune the algorithm for a selected type of data, by providing it with a few samples. The result of the training is stored in a file called "dictionary", which can be loaded before compression and decompression. Using this dictionary, the compression ratio achievable on small data improves dramatically:
Dictionary works if there is some correlation in a family of small data (there is no _universal dictionary_).
Hence, deploying one dictionary per type of data will provide the greatest benefits. Dictionary gains are mostly effective in the first few KB. Then, the compression algorithm will rely more and more on previously decoded content to compress the rest of the file.
Zstandard is currently deployed within Facebook. It is used daily to compress and decompress very large amounts of data in multiple formats and use cases.
