Research On Key Technologies For General-Purpose Computing On GPU In The Virtualization Environment

With the maturation of computer hardware and software development environment, GPUcomputing applications gradually evolved from graphic computing to general-purpose computingfield, and the scientific research examples of huge amounts of data computing carried on theplatforms of GPU-based high performance computing are numerous. While virtualization is a hottopic in today’s computer industry, the academic study in the junction of virtualization and GPUgeneral computing is still in its infancy. This paper focuses on the study of the key technologies usingthe GPU for general purpose computing in virtualization environment.In this paper, it makes a detailed analysis on virtualization technology and GPU computing and asummary on the existing GPU virtualization solutions. On the basis of GPU virtualization solutionand GPU resource scheduling algorithm, it brings up a way to improve the existing GPU schedulingalgorithm, which aims at reducing the turnaround time. By setting the value of a comprehensive loadevaluation it realizes the load balancing. Take the GPU features, the size and complexity of the taskinto the consideration of load evaluation to achieve a more fine-grained and more accurate loadevaluation. For large-scale computing programs, this paper designs a multi-GPU collaborativecomputing technology based on OpenMP in a virtual environment, which divides different tasksaccording to the different types and scales of tasks, and makes experimental verification ofrepresentative scientific computing examples such as classical matrix calculation and discrete Fouriertransformation of high complexity. Experimental results show that with the increase of the calculationscale, it can achieve a speed-up ratio close to the number of GPU.In order to reduce the GPU general computational overhead of performance due to thevirtualization itself, the paper makes a summary of the existing optimization solutions of inter-domainvirtual machine communication, and then get the optimal method of communication which has theleast affection in CUDA in a specific virtualization platform. Meanwhile, in the exchange mode ofGPU data it designed a method of using two data transmission ways to test the GPU synergisticcalculation, and at last get the main factors which affect multiple GPU collaborative computingefficiency via the comparision of theoretical calculating and actual measuring.