Adaptive Scheduling Policies Of Virtualized GPU Resource In Cloud Gaming

Cloud gaming renders the game on the Graphic Processing Unit (GPU)of the cloud and streams the running results over the network so thatclients can play high-end games without the support of the latest hardware.As the virtualization technology for GPUs develops and matures, cloudgaming has become an emerging application among cloud services.However, it is a big challenge to share GPU resources among virtualmachines in cloud gaming centers. Firstly, cloud gaming service providerstend to allocate one GPU exclusively for one cloud game. Hence there is agreat waste of GPU resources in cloud gaming centers. Secondly, lackingthe efficient sharing of the GPU resources in the virtualized environment,it cannot guarantee the Service Level Agreement (SLA) requirements foreach cloud game to concurrently run multiple games on one GPU. Inaddition, the performance of cloud games is inevitably undermined byvarious runtime uncertainties, which makes it harder to schedule GPUresources in cloud gaming.To address the three aforementioned challenges, this paper proposes alightweight framework that can efficiently schedule GPU resources in thevirtualized environment. In order to prove the practicality andeffectiveness of the proposed framework, three feedback control basedalgorithms are integrated within the framework. The framework leveragesthe technology of application programming interfaces (API) interception tointerpose the requests of GPU resources from the guest operating system(OS) to the host OS in the virtualized environment. In this way, thescheduling algorithms are intercepted into the request processes by theframework and thus the framework is able to schedule virtualized GPU resources for cloud games. Benefitting from API interception, none of theguest or host OS, the device drivers and virtual machine monitors requiremodifications to employ the proposed framework.A set of APIs are provided by the framework to make it be able tohost a variety of scheduling algorithms. The three adaptive algorithms,namely SLA-aware (SA) scheduling, Fair SLA-aware (FSA) schedulingand Enhanced SLA-aware (ESA) scheduling, are designed andimplemented by these APIs. They have different design goals. Featuringfeedback control, the algorithms can mitigate the impact of the runtimeuncertainties on the performance of the cloud games. They can alsoguarantee the SLA requirements of all the games and maximize the usageof GPU resources.The experiment results demonstrate that the framework as well as thethree algorithms are adapted to the cloud gaming scenarios. On one serverof the cloud gaming centers, they can schedule GPU resources amongvarious cloud games at the desired level. The performance overhead of theframework and the algorithms is limited to5-12%.