Parallel batch queries on dynamic trees: algorithms and experiments

TL;DR

This paper improves batch parallel dynamic tree algorithms, implements a versatile system supporting various queries, and demonstrates its efficiency and robustness through extensive experiments on different forest structures.

Contribution

It introduces generalizations of batch dynamic trees supporting arbitrary degrees and complex queries, along with the first comprehensive implementation and experimental evaluation.

Findings

Achieved good speedup in experiments

Supported a broad set of queries efficiently

Demonstrated robustness across different forest characteristics

Abstract

Dynamic tree data structures maintain a forest while supporting insertion and deletion of edges and a broad set of queries in $O(\log n)$ time per operation. Such data structures are at the core of many modern algorithms. Recent work has extended dynamic trees so as to support batches of updates or queries so as to run in parallel, and these batch parallel dynamic trees are now used in several parallel algorithms. In this work we describe improvements to batch parallel dynamic trees, describe an implementation that incorporates these improvements, and experiments using it. The improvements includes generalizing prior work on RC (rake compress) trees to work with arbitrary degree while still supporting a rich set of queries, and describing how to support batch subtree queries, path queries, LCA queries, and nearest-marked-vertex queries in $O(k + k \log (1 + n/k))$ work and polylogarithmic span. Our implementation is the first general implementation of batch dynamic trees (supporting arbitrary degree and general queries). Our experiments include measuring the time to create the trees, varying batch sizes for updates and queries, and using the tree to implement incremental batch-parallel minimum spanning trees. To run the experiments we develop a forest generator that is parameterized to create distributions of trees of differing characteristics (e.g., degree, depth, and relative tree sizes). Our experiments show good speedup and that the algorithm performance is robust across forest characteristics.