A Study of Performances Considering Communications on a GPU Cluster

Abstract

Multiple processing nodes with GPUs connected by a high speed network organize a GPU cluster environment. It is expected to achieve very higher performance than the one of CPU-based cluster due to the drastic performance growth of the GPU. This paper focuses on the performance impact when the tasks are distributed by different assignments using MPI framework. Due to the difference between the local communication in a processing node and the remote one among the nodes, the performance is very likely to be changed Applying famous benchmarks available in the internet, this paper evaluates the performance impacts of different task assignments among the nodes on the 32-node Tesla-based GPU cluster.