The Research And Realization Of GPU Virtualization Approach In High-performance Cloud

With the platform edge of cloud computing platform, virtualization technology can integrate computing resources, which can achieve centralized management and provide users of computing services. High-performance cloud computing platform requires high-performance computing resources. In the high-performance cloud computing platform,GPU as a kind of high-performance resource, how to make it pooled and provide high-performance applications for users, while schedule and manage the platform resources effectively, has become the recent hot spot and difficultly in the industry.(1)After studying computing virtualization, memory virtualization, network virtualization, this paper analyzes the virtualization implementation way and architecture of GPU devices in the cloud platform. On the basis of Microsoft's Azure cloud, this paper designs the solution of GPU virtualization in the cloud computing platform.(2)On the basis of study of existing cloud computing platform architecture and implementation mechanism, this paper designs the structure and work mechanism of cloud computing platform with integration of GPU resources, this paper uses PBS to achieve the Computing resource virtualization, and uses the way through vCPU management vGPU to make vGPU and vCPU work together in the cloud platform.(3) In order to make vGPU have general purpose of parallel computing, the paper designs vCUDA implementation of virtual machine on cloud computing platform, which client virtual machine calls the privileged virtual machine to meet the time division multiplexing of GPU resources for different users. Through comparative analysis of experiments, proves the feasibility of the solution.(4) By the scheduling experiment, the utilization of resources in the platform with scheduling policy has significantly improved.