Research On A GPU Multi-channel Virtualization Method And Fair-scheduling Strategy

With the rapid development of cloud computing and big data,the ways of getting information for users are diversified,such as multi-screen independent display.Cloud computing and big data analysis processing rely on parallel computing protential of GPU(Graphics Processing Unit)virtual subsystem.Virtualization technology has also made considerable progress and development in recent years.In modern virtual computing environment,both local client and remote server need GPU virtualization to support graphical user hidden in virtual machines(VM);Existing GPU virtualization technologies are either unable to take full advantage of GPU's powerful 2D/3D hardware accelerated graphics rendering performance,hardware encoding/decoding or high-performance parallel computing potential,or it has not be considered that the internal resource of GPU domain are fairly allocated between VMs with different performance requirements of GPU application load.Based on Xen Hypervisor,this paper studies the GPU virtualization technology in virtual machine system,including GPU virtualization fairness method research of virtual multi-channel,and GPU encoding/decoding performance and its fairness under a virtualization environment.The main research contents are as follows:1)We propose a GPU multi-channel virtualization architecture(A Virtual Multi-Channel GPU Fair-Scheduling Method for Virtual Machines,VMCG),model the corresponding credit allocating and transferring mechanisms,and redesign the virtual multi-channel GPU fair-scheduling algorithm.VMCG provides a separate V-Channel for each guest VM(DomU)that competes with other VMs for the same physical GPU resources,and each DomU submits command request blocks to its respective V-Channel according to the corresponding DomU ID.Through the virtual multi-channel GPU fair-scheduling algorithm,not only do multiple DomUs make full use of native GPU hardware acceleration,but the fairness of GPU resource allocation can also be significantly improved during GPU-intensive workloads from multiple DomUs running on the same host.Experimental results show that,for 2D/3D graphics applications,the performance is close to 96%of the native GPUs,performance is improved approximately 500%for parallel computing applications,and GPU resource-allocation fairness is improved approximately 80%for the hybrid application.2)We propose a scheme of GPU hardware encoding/decoding and fair sharing for a multi-virtual environment.In DomUs,we use FFmpeg to optimize video encoding/decoding processing,because of serving codec audio and video streaming,except for native graphic driver,native audio driver also needs to be run in DomUs;in DomO,we extend Meditor module and provide an independent V-Channel for each DomU to deduce the interference between DomUs.We also design the fair-scheduling algorithm for GPU hardware encoding/decoding.This scheme can achieve GPU hardware encoding/decoding in multiple virtual machines and its fairness.The experimental results show that G/C is approximately 2.2,1.4,respectively,for software encoding and decoding when running the same video application.After enable hardware acceleration,the G/C is 5.2,2.4,respectively.The performance is increased by approximately 145%.Its fairness also be increased by about 35%,and it also has a good stability.