A Load Balancing Strategy for Monte Carlo Method in PageRank Problem

Abstract

PageRank algorithm is key component of a wide range of applications. Former study has demonstrated that PageRank problem can be effectively solved through Monte Carlo method. In this paper, we focus on efficiently parallel implementing Monte Carlo method for PageRank algorithm based on GPU. Aiming at GPU, a parallel implementation must consider instruction divergence on the single instruction multiple data (SIMD) compute units. Due to the fact that low-discrepancy sequences are determined sequences, we adopt the low-discrepancy sequences to simulate the random walks in PageRank computations in our load balancing strategy. Furthermore, we allocate each thread of a block to compute a random walk of each vertex with a same low-discrepancy sequence. As a result, no idle thread exists in the PageRank computations and warp execution efficiency is up to 99%. Moreover, our strategy loads the low-discrepancy sequences into shared memory to reduce the data fetch cost. The results of experiments show that our strategy can provide high efficiency for Monte Carlo method in PageRank problem in GPGPU environment.