Efficient PageRank on GPU Clusters

概要

論文の詳細を見る
In this work, we report scalability of PageRank on multi-GPU clusters. Our target GPU clusters may contain more than one GPU accelerator per node. Iterative solvers for irregular sparse problems poorly scale with increasing number of processors because of load imbalance problem and network bottleneck. GPU computing units are too fast, for which network performance remains too low. Even the latest network hardware cannot provide bandwidth appropriate for high performance GPUs. In our previous work, we have introduced several implementation techniques and algorithms required for scalable sparse iterative solvers on multi-GPU extended clusters and evaluated those techniques on a Conjugate Gradient solver5),6). In this work, we present the GPU cluster performance evaluation of another important iterative method, PageRank. For GPU implementation of PageRank, although we are inspired by the techniques that are presented in [6], we cannot use them as is, since PageRank data has very different characteristics than Krylov Method data. Our PageRank implementation on GPUs is based on our work on an efficient CPU cluster algorithm7). In our experiments, we observe that PageRank achieves better scalability because of the enough data size to saturate GPUs. We observe scalability up to a hundred GPUs for PageRank, being at the same time almost 10 times faster than the CPU cluster implementation with the same number of CPU cores.
2010-12-09