I was looking for cloud gpgpu providers when I came across the new instance type ‘cluster GPU’ in US East (Virginia) zone on Amazon EC2. There are other vendors, but we need inexpensive local setups to build solutions that can eventually run in the cloud. Amazon seem to have exposed a real device per such VM and the solution looks inelastic in nature. It has been possible to expose GPUs to virtual machines on bare metal virtualisation technologies for some time, but sharing a single GPU among multiple VMs, or multiple GPUs shared by multiple VMs in a symmetric manner hasn’t got much attention from cloud iaas providers yet.
Project gvirtus (moved here) by UniParthenope Open Source Lab is an attempt to virtualise gpgpu and looks good architecturally. It was initially created at University of Naples Parthenope to enable virtualised cuda/opencl in qemu-kvm enabled virtual machines. Early outcome of a technology investigation allowed turning the framework into a general purpose virtualization of services.
If you like the basic theory (introduction and architecture) and are interested in contributing, Raffaele Montella has now moved the gvirtus project to Google Code for community development. I believe this project has the potential to bring inexpensive cloud gpu computing to the masses.
