GPGPU/CUDA in KVM or Xen

Question

We would like to enable direct access to CUDA-enabled GPUs within either KVM or Xen (or Citrix Xenserver). Specifically, we want to use PCI passthrough to access an NVIDIA Tesla S2050 unit, which is composed of 4 Tesla C2050 devices. I'm aware that Xenserver 5.6 supports SLI Multi-OS-enabled GPUs, but my understanding is that the C2050s are not SLI Multi-OS devices. I would prefer to avoid RPC-like solutions due to performance and maintenance issues that I've experienced.

Amazon EC2 offers Xen VMs with Tesla M2050 GPUs, which I believe are quite similar to our C2050s, so it seems like what I would like to do should be possible. Has anyone successfully passed a non-SLI Multi-OS GPU to Linux VM from either KVM or Xen and successfully run a CUDA application? If so, what virtualization solution did you use and what GPU was used?

I know it's not direct answer, but I've tried using a 200 series Nvidia with ESXi 4.1 and got nowhere. Even with the passthrough set up to the correct PCI addresses, the guests were not able to access the devices with the access CUDA programs want. — Marcin, May 13 '11 at 22:09

score 1 · Answer 1 · edited Aug 12 '12 at 22:03

Does XenServer Boston solve this issue?

http://blogs.citrix.com/2011/05/18/xenserver-boston-beta-program/

With the Boston release, a physical GPU can be assigned to a VM so the applications running in the guest can leverage GPU instructions (“GPU pass-thru”). This provides significant TCO benefits for the XenDesktop HDX 3D Pro technology used for delivery of CAD and other graphical applications via virtual desktops

score 1 · Answer 2 · edited Nov 14 '12 at 23:24

1

It looks like you need nVidia VGX boards (Kepler) and Xen/Citrix.

edited Nov 14 '12 at 23:24

Renan

16,976
8
69
88

answered Nov 14 '12 at 23:02

Richard Żak

111
3

GPGPU/CUDA in KVM or Xen

2 Answers2

Linked