StreamMAP: Automatic Task Assignment System on GPU Cluster

Abstract

GPU Clusters which use General-Purpose GPUs (GPGPUs) as accelerators are becoming more and more popular in high performance computing area. Currently the mainly used programming model for GPU cluster is hybrid MPI/CUDA. However, when using this model, programmers tend to need detailed knowledge of the hardware resources, which makes the program more complicated and less portable. In this paper, we present StreamMAP, an automatic task assignment system on GPU Clusters. The main contributions of StreamMAP are (1) It provides powerful yet concise language extension suitable to describe the computing resource demands of cluster tasks. (2) It maintains resource information and implements automatic task assignment for GPU Cluster. Experiments show that StreamMAP provides programmability, portability and performance gains.