Direct access to GPGPU do-able?

[Carsten Aulbert <carsten.aulbert@aei.mpg.de>]

Hi,

I'll start with a one-off question here, so please cc me on the reply. 

We are running a largish cluster and are currently buying GPGPU systems (Tesla 
and soon Fermi based). We will have at least 2 possibly 4 of these cards per 
box and have the problem that some codes need different CUDA kernel drivers to 
run. As these boxes have 4 CPU cores, 12 GB of memory and CPU-VT support we 
thought that this might be solvable by creating (para-) virtualized guests on 
the boxes and passing one GPGPU device into a guest at a time. In there we 
then can run any kernel/driver combo necessary.

But since my current virtualization experience only stretches to OpenVZ and 
VirtualBox (tinkering with Xen a couple of years back), I don't know if KVM is 
the right approach here. We need something which we can automatically set-up 
via CLI, i.e. starting and stopping the guests need to be fully automatic, we 
don't need a graphical environment within the guests, just plain text is good 
enough.

What do you think, is looking at KVM the right choice for this? Can we pass a 
device directly into a guest?

Cheers

Carsten

-- 
Dr. Carsten Aulbert - Max Planck Institute for Gravitational Physics
Callinstrasse 38, 30167 Hannover, Germany
Phone/Fax: +49 511 762-17185 / -17193
http://www.top500.org/system/9234 | http://www.top500.org/connfam/6/list/3