uSAP: An Ultra-Fast Stochastic Graph Partitioner

Stochastic graph partitioning (SGP) plays a crucial role in many real-world applications, such as social network analysis and recommendation systems. Unlike the typical combinatorial graph partitioning problem, SGP presents unique computational difficulties due to time-consuming sampling processes. To address this challenge, the recent HPEC launched the Stochastic Graph Partitioning Challenge (SGPC) to seek novel solutions from the high-performance computing community. Despite many SGP algorithms over the last few years, their speed-ups are not remarkable because of various algorithm limitations. Consequently, we propose uSAP, an ultra-fast stochastic graph partitioner to largely enhance SGP performance. uSAP introduces a novel strongly connected component-based initial block merging strategy to reduce the number of partitioning iterations significantly. To further improve the runtime and memory performance, uSAP adopts a dynamic batch parallel nodal block assignment algorithm and a dynamic matrix representation. We have evaluated uSAP on the 2022 official HPEC SGPC benchmarks. The results demonstrate the promising performance of uSAP on graphs of different sizes and complexities. For example, uSAP achieves 129.4x speed-up over the latest champion on a graph of 50K nodes.