Locality-based relaxation: An efficient method for GPU-based computation of shortest paths

This paper presents a novel parallel algorithm for solving the Single-Source Shortest Path (SSSP) problem on GPUs. The proposed algorithm is based on the idea of locality-based relaxation, where instead of updating just the distance of a single vertex v, we update the distances of v’s neighboring vertices up to k steps. The proposed algorithm also implements a communication-efficient method (in the CUDA programming model) that minimizes the number of kernel launches, the number of atomic operations and the frequency of CPU-GPU communication without any need for thread synchronization. This is a significant contribution as most existing methods often minimize one at the expense of another. Our experimental results demonstrate that our approach outperforms most existing methods on real-world road networks of up to 6.3 million vertices and 15 million arcs (on weaker GPUs)