The Research Of Virtualization Of Parallel Computing Based On Multi-GPGPU

GPU has significant advantages compared to CPU on computer power and energy consumption to be widely used in high performance computing. Virtualization technology is one of the main technologies in cloud computing, which can make multiple virtual machines share the GPU devices in the cluster transparently by shielding the hardware infrastructure. Therefore, it can reduce the resources cost and improve the resource utilization. At present, virtualization based General purpose GPU technology is still in the research stage. Resource sharing solutions are generally lacked of effective support for GPU in the virtual environment.The general computing framework CUDA is taken as the research object in this paper. We design a solution of virtualization of parallel computing based on multi-GPGPU. In this way, the solution can schedule the tasks dynamically and process the multi tasks concurrently. The specific work includes the following aspects:First, we introduce the GPU into the virtual machine by intercepting the library dynamically and design the dynamic allocation and management architecture based on multi GPU computing resources. The architecture includes virtual user layer, virtual resource management layer and virtual computing resource service layer. The general calculation adaptation problem in virtual environment is resolved and it can share the GPU resources among multiple computing nodes.Second, the solution of multi-GPU parallel computing under virtualization is proposed for the large-scale computing tasks. The multi-threads and flow mode are supported to implement the multi GPU cooperative computing. The multi-level storage structure for GPU, transmission, communication and other aspects are analyzed to accelerate the program. We implement the multi GPU cooperative computing for the loosely coupled interaction pattern of data(such as Monte Carlo method) and tight coupled interaction model(such as QFT algorithm).Third, the Dynamic and Multi-Load Status algorithm for GPU(DMLS-GPU)has been proposed for estimating GPU computing capacity, which combines the load value with the GPU equipment hardware capacity and the task characteristic. And it resolves the dynamic evaluation of GPU computing capacity in the virtualization solution. The experiment shows that multiple CUDA programs can be performed concurrently on one or more GPU devices under virtualization. And we verify the good extension ability and the high efficiency of the virtualization solution in this paper.This paper is aimed at the challenges and constraints in GPU general purpose computing under virtualization. Multi-tasks share the GPU resources and multi-GPU parallel computing is studied under virtualization in order to further expanding their application space.