Revisiting Data Compression in Column-Stores

TL;DR

This paper reevaluates data compression techniques in disk-based column-stores, considering modern hardware advancements and new algorithms, to determine optimal compression strategies for query performance and storage efficiency.

Contribution

It provides an experimental analysis of heavy-weight versus light-weight compression schemes and assesses the impact of SIMD-based decompression in modern disk-based column-stores.

Findings

Heavy-weight compression schemes are still generally unsuitable for disk-based column-stores.

New light-weight algorithms outperform older ones in modern hardware contexts.

SIMD-based decompression offers performance benefits in disk-based systems.

Abstract

Data compression is widely used in contemporary column-oriented DBMSes to lower space usage and to speed up query processing. Pioneering systems have introduced compression to tackle the disk bandwidth bottleneck by trading CPU processing power for it. The main issue of this is a trade-off between the compression ratio and the decompression CPU cost. Existing results state that light-weight compression with small decompression costs outperforms heavy-weight compression schemes in column-stores. However, since the time these results were obtained, CPU, RAM, and disk performance have advanced considerably. Moreover, novel compression algorithms have emerged. In this paper, we revisit the problem of compression in disk-based column-stores. More precisely, we study the I/O-RAM compression scheme which implies that there are two types of pages of different size: disk pages (compressed) and in-memory pages (uncompressed). In this scheme, the buffer manager is responsible for decompressing pages as soon as they arrive from disk. This scheme is rather popular as it is easy to implement: several modern column and row-stores use it. We pose and address the following research questions: 1) Are heavy-weight compression schemes still inappropriate for disk-based column-stores?, 2) Are new light-weight compression algorithms better than the old ones?, 3) Is there a need for SIMD-employing decompression algorithms in case of a disk-based system? We study these questions experimentally using a columnar query engine and Star Schema Benchmark.