Label Propagation-Based Parallel Graph Partitioning for Large-Scale Graph Data

Abstract

The increasing importance of graph data in various fields requires large-scale graph data to be processed efficiently. Furthermore, well-balanced graph partitioning is a vital component of parallel/distributed graph processing. The goal of graph partitioning is to obtain a well-balanced graph topology, where the size of each partition is balanced while the number of edge cuts is reduced. Moreover, a graph-partitioning algorithm should achieve high performance and scalability. In this study, we present a novel graph-partitioning algorithm that ensures a high edge cutting quality and excellent parallel processing performance. We apply formulas based on the label propagation algorithm to improve the quality of edge cuts and achieve fast convergence. In our approach, the necessity of applying the label propagation process for all vertices is removed, and the process is applied only for candidate vertices based on a score metric. Our proposed algorithm introduces a stabilization phase in which remote and highly connected vertices are relocated to prevent the algorithm from becoming trapped in local optima. Comparison results show that a prototype based on the proposed algorithm outperforms well-known parallel graph-partitioning frameworks in terms of speed and balance.