Towards GPGPU Assisted Computing in Virtualized Environments
Institute of Distributed Systems, Ulm University, James-Franck-Ring O-27/3210, 89069 Ulm, Germany
EURO-PAR 2010, Parallel Processing Workshops, Lecture Notes in Computer Science, 2011, Volume 6586/2011, 649-656
General Purpose Computation on Graphics Processing Units (GPGPU) makes it possible to use the massive computing power of modern graphics cards for generic high-performance computing. However, the new virtualization technologies will typically not support high-performance graphics cards and as a consequence GPGPU resources can not be used in typical virtualization setups. In this paper we present an approach to introduce accelerated 3D graphics support as well as GPGPU facilities into virtualized environments. We present our proof-of-concept "VirtGL" and discuss architectural considerations and a lean, straight-forward way of implementation. We then give an outlook on how this approach can easily be adopted for virtualizing GPGPU APIs like CUDA, Stream, OpenCL and alike.
September 8, 2011 by hgpu