Practical Parallel Block Tree Construction: First Results

TL;DR

This paper introduces efficient parallel algorithms for constructing block trees, significantly reducing memory usage and construction time, making compressed text indexing more practical for large datasets.

Contribution

The paper presents the first practical parallel algorithms for block tree construction, achieving faster speeds and lower memory requirements than existing methods.

Findings

Parallel algorithms match fastest single-core construction speeds

Achieve up to four times faster construction with 64 cores

Require an order of magnitude less memory

Abstract

The block tree [Belazzougui et al., J. Comput. Syst. Sci. '21] is a compressed representation of a length-$n$ text that supports access, rank, and select queries while requiring only $O(z\log\frac{n}{z})$ words of space, where $z$ is the number of Lempel-Ziv factors of the text. In other words, its space-requirements are asymptotically similar to those of the compressed text. In practice, block trees offer comparable query performance to state-of-the-art compressed rank and select indices. However, their construction is significantly slower. Additionally, the fastest construction algorithms require a significant amount of working memory. To address this issue, we propose fast and lightweight parallel algorithms for the efficient construction of block trees. Our algorithm achieves similar speed than the currently fastest construction algorithm on one core and is up to four times faster using 64 cores. It achieves all that while requiring an order of magnitude less memory. As result of independent interest, we present a data parallel algorithm for Karp-Rabin fingerprint computation.