CUDA Virtualization on XEN

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,
I have a class project in which we wish to virtualize CUDA on XEN. I tried looking up on various forums, and people have commented - how they can have NVIDIA driver ported to XEN, without any patch.
Step by step guide NVIDIA on XEN :
[url=“http://www.nvnews.net/vbulletin/showthread.php?t=122900&highlight=xen”]http://www.nvnews.net/vbulletin/showthread...p;highlight=xen[/url]
I did not check if this works, since I am unable to install NVIDIA driver on a normal kernel.

Does this mean, that we will not have to write driver for Dom0 of Xen, but might have to write a front end driver for DomU?
Also, for virtualizing CUDA, we will need the source code of the CUDA driver, Does NVIDIA have a repository from where one could download the source code of Linux drivers?

Thanks.

2

HI!

Maybe this answer is too late, but, anyway, we have the solution: gVirtuS (http://osl.uniparthenope.it/projects/gvirtus/)

There are no officially supported NVIDIA drivers working on XEN, so you have to walk around in this way:

  1. Create a virtual machine running on Xen, use and connect the real device (NVIDIA card) to the virtual machine bonding the PCI slot. This virtual machine can be as small as you can. You have just to run on it a small footprint linux distro with the gVirtuS backend.

  2. Then create as much as you can virtual machines on XEN and run on it the gVirtus frontend.

  3. Finally use the tcp communicator between front and backend setting up Xen Loop

  4. If you want, you can contribute in gVirtuS development writing a better shared memory based communicator working on XEN.

Keep me up to date,

Raffaele