Efficient parallelism in Breadth-First Search: A comprehensive analysis and implementation

Abstract

The Breadth-First Search (BFS) entails a systematic traversal of a given graph, G = (V, E), layer by layer, starting from a specific vertex. Recognized as a cornerstone methodology for graph exploration, the importance of BFS has skyrocketed, especially with the increasing demands of graph-based data processing. However, as the vertex count expands, traditional serial implementations reveal their limitations, faltering in terms of time and space efficiency. This paper aims to contrast the efficiencies of standard BFS with its parallelized iteration. Introducing a shared-memory model of level-synchronous parallel BFS, the approach integrates optimizations to navigate the challenges posed by implicit barriers and critical sections. Employing the Graph500 benchmark, this parallel methodology is meticulously evaluated, highlighting the speedup concerning various thread counts. Initial findings unveil a compelling pattern: speedup generally correlates positively with the number of active threads. However, if the thread count breaches the system's inherent capacity, the speedup hits a plateau, showing only marginal fluctuations without significant increases. These statistical revelations not only vouch for the advantages of BFS parallelization but also emphasize a critical insight: judiciously increasing thread count, up to a system-specified limit, can yield peak efficiency.