FUSED-PAGERANK: Loop-Fusion based Approximate PageRank

Abstract

PageRank is a graph centrality metric that gives the importance of each node in a given graph. The PageRank algorithm provides important insights to understand the behavior of nodes through the connections they form with other nodes. It is an iterative algorithm that ranks the nodes in each iteration until all the node values converge. The PageRank algorithm is implemented using sparse storage format, which results in irregular memory accesses in the code. This key feature inhibits optimizations to improve its performance, and makes optimizing the PageRank algorithm a non-trivial problem. In this work we improve the performance of PageRank algorithm by reducing its irregular memory accesses. In this paper, we propose FUSED-PAGERANK algorithm, a compiler optimization oriented approximate technique that reduces the number of irregular memory accesses in the PageRank algorithm, improving its locality while making the convergence of the algorithm faster with better accuracy in results. In particular, we propose an approximate PageRank algorithm using Loop-Fusion. We believe that ours is the first work that formally applies traditional compiler optimization techniques for irregular memory access in the PageRank algorithm. We have verified our method by performing experiments on a variety of datasets: LAW graphs, SNAP datasets and synthesized datasets. On these benchmarks, we have achieved a maximum speedup (vs. -O3 optimization) of 2.05X, 2.23X, 1.74X with sequential version, and ~4.4X, ~2.61X, ~4.22X with parallel version of FUSED-PAGERANK algorithm in comparison with Edge-centric version of PageRank algorithm.