Computing Betweenness Centrality for Small World Networks on a GPU

Although a graphics processing unit (GPU) is a specialized device tailored primarily for compute-intensive, highly dataparallel computations; significant acceleration can be achieved on memory-intensive graph algorithms as well. In this work, we investigate the performance of a graph algorithm for computing vertex betweenness centrality for small world networks on 2 NVIDIA Tesla and Fermi GPUs and compare it to a parallel open source implementation on an Intel multicore CPU. For the test instances considered the betweenness computation on GPU was accelerated by as much as 19.68× compared to single thread CPU performance and more than 2× compared to multithread CPU performance using 16 OpenMP threads. Introduction Network analysis is an active area of research with applications in variety of domains such as social networks, protein interaction networks, computer security and disease spread. These real-world networks though highly unrelated display common features such as a small diameter, power law degree distribution and community structure, or in other words a small world topology. They are often very large in size with number of vertices and edges varying from millions to billions. These graphs are sparse and their analysis tends to be highly memory intensive. They have a large memory footprint, and a significant number of noncontiguous memory accesses to global data structures with low degree of spatial and temporal locality. Compared to other workloads most graph algorithms have little computation to hide latency to memory access. Betweenness centrality is a key metric that is used to identify important actors in a network. It is a popular graph analysis technique based on shortest path enumeration. For a graph with vertices and edges, the betweenness centrality of vertex is defined as,