| With Compute Unified Device Architecture, Graphic Processing Unit (GPU) can dealgeneral-purpose computing tasks efficiently, and the GPU programs can get several times,even ten times speedup compared with CPU programs. At the same time, virtualizationtechnology becomes more and more mature. The combination of virtualization technologyand GPU computing is favored in the area of high performance computing. However,there is no GPU computing management system based virtual machine currently, or thecurrent cluster systems suffer from the following defects: the virtual machines can’tsupport CUDA computing, or the cluster system based on physical machines can’t managethe resources efficiently and easily.For GPU computing based on virtualization technology, the GPU computingresources management system based on virtual machines is proposed, known asVMGPURMS. VMGPURMS supports GPU computing in virtual machines and usesCPU-GPU hybrid scheduling strategies to improve system utilization. VMGPURMS usesshared memory model for data communication between privileged domain and virtualmachine to reduce the overhead of data communication; CPU-GPU hybrid schedulingstrategies are: Least resource strategy is used to avoid unnecessary resource costs, whenthe number of GPUs is greater than CPUs; Resource reservation strategy is adopted toprepare corresponding CPU and GPU resources for GPU jobs, when both GPU and CPUresources are necessary for GPU jobs; Once CPU jobs occupy GPU resources, GPU jobsseize strategy is used to shift CPU jobs into other node, so as to release GPU resources toGPU jobs.The system implements collaborative resources management and scheduling betweenprocessor and graphics resources based on virtual machines. The results of tests show thatthe performance of GPU computing in the virtual machine is only lost15%(comparedwith physical environment). After using the proposed CPU-GPU hybrid schedulingstrategies, the system utilization can be increased about17%. |
PDF Full Text Request