A comparison of the performance of graph breadth-first search algorithms based on OPENMP and CUDA

Graphs, as fundamental data structures, are indispensable for delineating intricate interrelations among entities, exhibiting utility across diverse domains within the realm of the information society. Traversal algorithms, including the pivotal breadth-first search, are keystones in the realm of graph theory, constituting essential elements within broader algorithmic frameworks. With the expansion of graph scale, the imperative for expedient processing of graph data becomes increasingly pronounced. This paper aims to meticulously examine the execution efficiency of three distinct breadth-first traversal methodologies across varying node magnitudes: the traditional queue-based sequential approach, CPU parallelization leveraging OPENMP and GPU parallelization harnessing CUDA. Each method possesses its distinctive array of advantages and disadvantages contingent upon the scale of nodes. Moreover, this study aims to delve into the nuanced effect of thread count on the efficacy of OPENMP-based CPU parallel traversal, elucidating an inversely proportional relationship in scenarios with fewer nodes. Finally, we investigate the influence of graph sparsity on the traversal methodologies, acknowledging its pivotal role in connectivity and traversal time.