Evaluating GPU Performance for Deep Learning Workloads in Virtualized Environment

Abstract

Deep Learning (DL) is the fastest growing high performance data center class workload today. Deep learning algorithms render themselves well to taking advantage of GPU parallelism, therefore GPGPU acceleration is a mainstay of the DL computing infrastructure. In this paper we evaluate virtualized GPU performance based on training of state-of-the art deep learning models. We find that there is a correlation between the amount of I/O traffic generated in the deep learning training workload and the efficiency of GPGPU performance in virtualized environments. We show that one can achieve high efficiency when using GPGPUs in virtualized and networkattached multi-GPU environments to perform highly computeintensive workloads.