Research On GPU Cluster Resource Virtualization Theory And Cloud Sharing Technology

Photorealistic 3D scene rendering technology is maturing and has been applied to many different areas.However,with the technology becoming mature,People's requirements for the scene are also more complicated,and ordinary PC no longer satisfy real-time scene rendering.Therefore,3D scene rendering in the cloud is proposed to solve the requirement of computing capability in scene rendering through the powerful computing power in the cloud.That is why GPU sharing has become one of the hot spot in the field of computer graphics rendering.This thesis is about GPU cluster virtualization and cloud sharing framework,and the main contents are as follows:First,introduced the GPU virtualization and GPU cluster sharing scheduling technology,including virtualization technology,as well as some research progress of load balancing strategy in GPU scheduling.Second,a resource scheduling strategy based on GPU virtual cluster is proposed.This scheduling strategy is by mapping the GPU resources and drawing tasks into hash space,then use hash algorithm choosing compute node for drawing tasks.Trough this scheduling strategy solve the problem of server downtime well,At the same time,through the optimization of the algorithm,the problem of uneven distribution of tasks is well solved.Third,due to computing resources are allocated evenly in the process of GPU task allocation,so it is also required to divide the scene into many small scenes uniformly when the scene is divided by the client.This thesis use algorithm base on pixel recursive to divide the scene.Finally,through the implementation of the above algorithm,this thesis implements a cluster resource sharing system based on GPU virtualized.