On GPU pass-through performance for cloud gaming: Experiments and analysis

Abstract

Cloud Gaming renders interactive gaming applications remotely in the cloud and streams the scenes back to the local console over the Internet. Virtualization plays a key role in modern cloud computing platforms, allowing multiple users and applications to share a physical machine while maintaining isolation and performance guarantees. Yet the Graphical Processing Unit (GPU), which advanced game engines heavily rely upon, is known to be difficult to virtualize. Recent advances have enabled virtual machines to directly access physical GPUs and exploit their hardware's acceleration. This paper presents a experimental study on the performance of real world gaming applications as well as ray-tracing applications with GPUs. Despite the fact that the VMs are accelerated with dedicated physical GPUs, we find that the gaming applications perform poorly when virtualized, as compared to non-virtualized bare-metal base-line. For example, experiments with the Unigine gaming benchmark run at 85 FPS on our bare-metal hardware, however, when the same benchmark is run within a Xen or KVM based virtual machine the performance drops to less than 51 FPS. In contrast, ray-tracing application fares much better. Our detailed performance analysis using hardware profiling on KVM further reveals the memory bottleneck in the pass through access, particularly for real-time gaming applications.