Efficient implementation of the Bellman-Ford algorithm on GPU

Single-source shortest path (SSSP) problem is a common algorithm in graph analysis. Bellman-Ford is a primitive algorithm to solve the shortest path to the source node, which enables the detection of the negative-weighted cycle in a graph. Moreover, it is represents a class of parallel algorithms, the memory accesses and work distribution of which are both irregular and data-dependent. Recently, graphics processors have been used for implementing many algorithms, as well as an accelerator in supercomputers. Several SSSP algorithms have been proposed based on graphics processing units (GPUs), each of which could traverse a specific type of graph. In this paper, we accelerated the Bellman-Ford algorithm on GPU using CUDA, so that it could traverse dense and sparse graphs (regular and irregular) within the shortest time compared to the previous algorithms. According to the simulation results, the proposed implementation provided an average speed-up of 1.87× compared to most of the previous parallel implementation algorithms.