Resource Optimization Scheduling Based On GPU Virtualization

With the development of big data and cloud computing,application deployment on cloud computing platform has become the first choice for users.In some application fields that require the support of distributed GPU computing power,there are differences in scene-based requirements for GPU computing performance,computing resources,and API compatibility.Since different types of GPUs differ in chip architecture,driver and API compatibility,and virtualization methods,it is difficult for traditional GPU virtualization technologies to achieve complete compatibility with these heterogeneous GPU chips.At the same time,with the continuous development of GPU technology,the demand for GPU computing power in cloud computing data centers is also increasing in the fields of graphics and image rendering,3D simulation design,and high-performance computing.How to achieve compatibility with these heterogeneous GPU chips through a set of virtualization framework,and how to allocate and schedule these GPU resources in the cloud computing platform becomes extremely important.Therefore,this paper firstly optimizes and improves the traditional GPU virtualization framework,and then proposes an iteratively optimized heuristic resource allocation algorithm from the perspective of fairness and efficiency,and proposes a virtual GPU resource for the problem of insufficient utilization of virtual GPU resources.Fine-grained scheduling strategy for GPU resources.Finally,experiments and analysis are carried out on the simulation platform.The main work and contributions of this thesis are as follows:(1)An IGVF(Improved Gvirtus Virtualization Framework)framework based on a general GPU virtualization framework Gvirtu S is proposed,which can manage multiple physical GPUs at the same time.The scheduler of the IGVF framework can manage and allocate the GPU resources of multiple physical machines.There is no strict requirements on the chip architecture and model of these physical GPUs so that it supports multiple virtualization technologies.The IGVF framework includes a GPU cloud computing power center and a GPU cloud control node,which can provide GPU computing power and resource scheduling instructions,respectively.Experimental results of performance comparison based on CUDASW++ software show that the improved IGVF framework in this paper enables virtual GPUs to have performance close to physical GPUs.(2)An iterative optimization heuristic resource allocation algorithm is proposed.On the fairness metric model,the quantization method and metric function of the FDS function family are improved,and the idea of approximating the optimal point through a predefined step size iteration in the subgradient method is incorporated,and a heuristic method is used to adjust the value of each iteration.step size.The algorithm can ensure the fairness and efficiency of virtual GPU resource allocation on the IGVF framework.The performance test is carried out on the Cloud Sim simulation platform.Compared with the classic adaptive algorithm v GASA,the fairness efficiency value of our algorithm is improved by 4.5%.(3)A fine-grained scheduling strategy for co-optimized virtual GPU resources is proposed.Using a hybrid time-or event-scheduling strategy successfully improves the traditional Credit scheduling strategy and solves the iusse of wasting idle time during task switching.A time-sharing sharing scheduling strategy to utilize the GPU idle time is added when tasks are suspended.Then,simulation experiments are carried out on the Cloud Sim framework to simulate resource scheduling on a large cloud computing platform.The performance comparison experiments were carried out in the case of the optimization method,and the results show that our strategy has certain advantages in load balancing,and the execution time of the algorithm is also significantly reduced when considering suboptimal scheduling.Finally,the practical application case of IGVF framework,fair and efficient allocation algorithm and collaboratively optimized Credit scheduling strategy on a government cloud platform is shown.