Towards Exploiting GPUs for Fast PageRank Computation of Large-Scale Networks

Abstract

The size of network (or graph) is increasing, and so, a fast algorithm for network analysis is more important than ever. PageRank-style algorithm is one of the most important and fundamental algorithms for network analysis. Meanwhile, the paradigm of micro-architecture design of computer processors has been shifted to on-chip multi-core CPUs, and furthermore, many-core GPUs. The current fastest PageRank method is the one based on multi-core CPUs. In contrast, there is lack of studies on using many-core GPUs for networks analysis including PageRank yet due to difficulty to develop a GPU algorithm that efficiently manipulates complex and irregular data structures like complex networks. This paper proposes novel fast parallel PageRank methods that exploit the massive parallelism of GPU for large-scale networks. More specifically, the paper proposes the node-centric method computing PageRank in terms of nodes and the edge-centric method computing PageRank in terms of edges. They efficiently compute PageRank based on a GPU with compact data structures and concise kernel functions. Through extensive experiments, the paper shows the proposed methods outperform the state-of-the-art method by up to about two times.