In this video from the GPU Technology Conference, John Paul Walters from the USC Information Sciences Institute presents: Achieving Near-Native GPU Performance in the Cloud.
In this session we describe how GPUs can be used within virtual environments with near-native performance. We begin by showing GPU performance across four hypervisors: VMWare ESXi, KVM, Xen, and LXC. After showing that performance characteristics of each platform, we extend the results to the multi-node case with nodes interconnected by QDR InfiniBand. We demonstrate multi-node GPU performance using GPUDirect-enabled MPI, achieving efficiencies of 97-99% of a non-virtualized system. Examples are drawn from signal processing, big data analytics, and molecular dynamics. The session will conclude with a discussion of the next steps in extending HPC to virtual environments, including our work with the OpenStack platform.”