Edge v. Node Parallelism for Graph Centrality Metrics

This chapter proposes an improved edge-parallel approach for computing centrality metrics that can also accelerate breadth-first search and all-pairs shortest path. For some graph algorithms such as computing centrality, breadth-first search, and even all-pairs shortest path, an edge-parallel approach improves graphical processing unit (GPU) throughput with better load balancing and less thread divergence on scale-free networks. The edge-parallel approach is less appropriate for grids, meshes, and other graphs with low-degree variance, and it performs less well on dense graphs with many edges. The edge-parallel approach requires more GPU memory than node-parallel, which could limit its application to larger graphs than the considered, for which one could investigate efficient inter-block synchronization techniques. It has been believed the edge-parallel approach would benefit most scale-free network applications and should be investigated on further graph algorithms such as max flow.