GPU Video Codec And Dynamic Graphic Memory Allocation Scheme In Virtual Environment

In today's society where computer science is rapidly developing,the way of information interaction is changing with each passing day,especially in the way of large amounts of data such as audio and video.The emergence of virtualization technology enables humans to maximize the use of expensive physical devices,especially the virtualization of GPU(Graphic Processing Unit)has become a hot topic of virtualization research.Although the current GPU virtualization technology has been able to support physical GPU-accelerated 2D/3D graphics rendering and high-performance parallel computing in virtual machines,there are some deficiencies and unconsidered places.For example,at present,there is no good solution for the video codec accelerated by the physical GPU in the virtual environment,which fails to exert the huge advantage of the GPU in data computing,and there is also room for optimization in the problem of GPU resource allocation.What's more,because of the allocation of GPU resources,there will be problems in how to efficiently allocate graphic virtual memory in a virtual environment.Based on the Xen virtualization architecture,this paper proposes how to implement the physical GPU accelerated codec method and the dynamic allocation algorithm of virtual graphic memory in the virtual environment.The main research content of this paper can be summarized as the following two points:(1)In this paper,we propose a GPU hardware video codec method in multiple virtual domain environments(VMH-codec).First,we design VMH-codec architecture by exploring video codec engine inside GPU and the pass-through of gVirt,which offloads the CPU overheads and enhances the performance of video codec.Then,especially aiming to avoiding too many lags in video codec,we redesign a video codec priority scheduling algorithm and fairness scheduling method between VMs,which maintains the good smoothness and stability for VMH-codec.The experimental results show that VMH-codec can realize physical GPU accelerated video encoding and decoding,solve the problem of video jam in virtual machine,and obtain good video playback effect.The fairness of GPU resource allocation and application performance are significantly improved.Compared with the gVir architectur e,the fairness is improved by about 35%,and the video codec performance is improved by about 40%.(2)This paper refers to the virtual machine memory allocation me thods such as ballooning technology and memory hotplug technology,and combines the address space ballooning mechanism used by gVirt architecture to design a new dynamic allocation algorithm for graphic memory.We collect the memory usage of each virtual machine,the change of the swap space and the incremental coefficien t of the GPU page cache hit ratio in a certain period of time,and submit the data to the global memory allocator in the host machine.Finally,the allocator calculates the exact graphic memory requirement of each virtual machine in the next time window.Through the virtual memory graphic dynamic allocation algorithm,not only multiple virtual machines can fully utilize limited graphic memory resources,each virtual machine can obtain corresponding size graphic memory according to its real needs for graphic memory resources,and achieve fair distribution of global graphic memory.The experimental results show that compared with gVirt's static memory allocation algorithm,the dynamic memory allocation algorithm improves performance by about 32% and maintains good stability.