UC Davis

Finding the shortest paths from a single source to all other vertices is a fundamental method used in a variety of higher-level graph algorithms. We present three parallel friendly and work-efficient methods to solve this Single-Source Shortest Paths (SSSP) problem: Workfront Sweep, Near-Far and Bucketing. These methods choose different approaches to balance the tradeoff between saving work and organizational overhead. 

In practice, all of these methods do much less work than traditional Bellman-Ford methods, while adding only a modest amount of extra work over serial methods. These methods are designed to have a sufficient parallel workload to fill modern massively-parallel machines, and select reorganizational schemes that map well to these architectures. We show that in general our Near-Far method has the highest performance on modern GPUs, outperforming other parallel methods.

We also explore a variety of parallel load-balanced graph traversal strategies and apply them towards our SSSP solver. Our work-saving methods always outperform a traditional GPU Bellman-Ford implementation, achieving rates up to 14x higher on low-degree graphs and 340x higher on scalefree graphs. We also see significant speedups (20–60x) when compared against a serial implementation on graphs with adequately high degree.