VGVM: Efficient GPU capabilities in virtual machines
Dimitrios Vasilas, Stefanos Gerangelos, Nectarios Koziris · 2016
Graphics Processing Units (GPUs) have become a powerful platform, that can provide significant performance benefits to data parallel applications. Graphic processors are being increasingly introduced as accelerators in high performance computing (HPC) systems due to the development of GPGPU (General-Purpose Computation on GPUs). Furthermore, virtualization technologies are gaining interest in these domains, due to their benefits on server consolidation as well as the isolation and ease of management they offer. There is thus a growing need to combine the benefits of both fields by providing heterogeneous resources, particularly GPUs, in virtual environments. In this paper we address the challenge of integrating GPGPU into virtualized environments. We propose VGVM, a mechanism that enables the execution of GPU accelerated applications within Virtual Machines (VMs). Our framework consists of two components: a user level library and a paravirtualized driver, which enables communication with the host's GPU driver. To validate our approach, we conduct experiments on a variety of GPU applications, focusing on the virtualization overhead and the scalability of our framework.