TSXor: A Simple Time Series Compression Algorithm

Developed @ ISTI CNR - HPC Lab (Pisa)

Full paper here

Code

• Introduction
• How it works
• Building the code
• Input data format
• Run
• Benchmarks
• Authors
• Notes

Introduction

TSXor, a simple yet effective encoder/decoder for time series that achieves high compression ratios and fast decompression speed. TSXor exploits the redundancy/similarity between close-in-time values through a window that acts as a cache.

How it works

Window

TSXor compares each v_n value with its preceding W ≤ 127 values, logically corresponding to the values seen in the time range [t_n−W , t_n−1]. The goal is to compress v_n relative to this “window” containing the previous W values. We distinguish between 3 cases, namely Reference, XOR, and Exception.

Reference

If v_n is equal to a value in the window, just output its position p in the window. Since the window contains at most 127 values, 1 byte suffices to write the position with the most significant bit always equal to 0.

If the window does not contain v_n, then we search for the value u in the window such that x = vn ⨁ u has the largest number of leading (LZ) and trailing (TZ) zeros bytes. Let p be the position of u in the window. We first write p + 128 using 1 byte. In this case the most significant bit will always be 1 because of sum, which allows us to distinguish this case from the Reference case.

XOR

If LZ +TZ ≥ 2, we output a byte where 4 bits are dedicated to TZ and the other 4 bits to the length (in bytes) of the segment of x between the leading and trailing zero bytes. We then write such middle bytes.

Exception

We output an exception code, i.e., the value 255 using 1 byte, followed by the plain double-precision representation of v_n using 8 bytes.

Builiding the code

The code has been tested both on Linux and MacOS.

No dependencies are needed.

Just clone this repo and execute:

make all

Input data format

The algorithm can process any .csv file containing numbers only. You need first to convert the .csv into a .bin file using the csv_to_bin utility as follows:

cd util

./csv_to_bin.o path/to/MY_DATASET.csv

Please note: the first column will be interpreted as the timestamp, the rest will be interpreted as values.

Run

Compression

To run a compression test of a .bin file, execute the following commands:

cd test

./compression.o path/to/MY_DATASET.bin

This will produce a file called compressed_data.tsx

Decompression

To decompress the file compressed_data.tsx, run the command:

./decompression.o

Benchmarks

The following tables show the comparison between TSXor with respect to Gorilla by Facebook and FPC by Burtscher and Ratanaworabhan. The experiments were run on an Ubuntu 18.04 machine with Intel i7-7700 CPU @ 3.60GHz.

Compression Speeds (MB/s)

Decompression Speeds (MB/s)

Compression Ratios

Authors

• Andrea Bruno
• Franco Maria Nardini
• Giulio Ermanno Pibiri
• Roberto Trani
• Rossano Venturini

Notes

This is a beta version. Use it at your own risk.