Machine Learning Using Virtualized GPUs in Cloud Environments

摘要

Using graphic processing units (GPU) to accelerate machine learning applications has become a focus of high performance computing (HPC) in recent years. In cloud environments, many different cloud-based GPU solutions have been introduced to seamlessly and securely use GPU resources without sacrificing their performance benefits. Among them are two main approaches: using direct pass-through technologies available on hypervisors and using virtual GPU technologies introduced by GPU vendors. In this paper, we present a performance study of these two GPU virtualization solutions for machine learning in the cloud. We evaluate the advantages and disadvantages of each solution and introduce new findings of their performance impact on machine learning applications in different real-world use-case scenarios. We also examine the benefits of virtual GPUs for machine learning alone and for machine learning applications running together with other GPU-based applications like 3D-graphics on the same server with multiple GPUs to better leverage computing resources. Based on our experimental results benchmarking machine learning applications developed with TensorFlow, we discuss the scaling from one to multiple GPUs and compare the performance between two virtual GPU solutions. Finally, we show that mixing machine learning and other GPU-based workloads can help to reduce combined execution time as compared to running these workloads sequentially.