Computing $k$-Bisimulations for Large Graphs: A Comparison and Efficiency Analysis

TL;DR

This paper presents a parallel BRS algorithm for computing $k$-bisimulations on large graphs, demonstrating superior performance over existing algorithms for large $k$ values across diverse datasets.

Contribution

It introduces a generic, efficient parallel BRS algorithm for $k$-bisimulation, outperforming existing methods on large-scale graphs and providing a comprehensive performance comparison.

Findings

BRS outperforms native bisimulation algorithms for $k extgreater=5$.

BRS is effective for large graphs with up to two billion edges.

Parallelization makes the bisimulation computation scalable and efficient.

Abstract

Summarizing graphs w.r.t. structural features is important to reduce the graph's size and make tasks like indexing, querying, and visualization feasible. Our generic parallel BRS algorithm efficiently summarizes large graphs w.r.t. a custom equivalence relation $\sim$ defined on the graph's vertices $V$. Moreover, the definition of $\sim$ can be chained $k\geq 1$ times, so the defined equivalence relation becomes a $k$-bisimulation. We evaluate the runtime and memory performance of the BRS algorithm for $k$-bisimulation with $k=1,\ldots,10$ against two algorithms found in the literature (a sequential algorithm due to Kaushik et al. and a parallel algorithm of Sch\"atzle et al.), which we implemented in the same software stack as BRS. We use five real-world and synthetic graph datasets containing 100 million to two billion edges. Our results show that the generic BRS algorithm outperforms the respective native bisimulation algorithms on all datasets for all $k\geq5$ and for smaller $k$ in some cases. The BRS implementations of the two bisimulation algorithms run almost as fast as each other. Thus, the BRS algorithm is an effective parallelization of the sequential Kaushik et al. bisimulation algorithm.