Good day, readers! I want to share with you some ideas on GPU virtualisation.

This topic is pretty new for me, however, since I had experience with CUDA, and some times was called 'The gpu master' by our prof, i decided to take this as a challenge for myself

First I looked for official NVIDIA virtualization support. It is indeed provided, for GRID family of devices.

So there are three ways left :

Use KVM with RedHat (wasn’t concerned too much since it is not free software)

Using Xen product, which has some doubtful results (according to results provided on forums)

Using quemu (alternative of KVM for Ubuntu with no RedHat kit) - currently in work.

This third method is based on pci passthrough using pci stub and then passing through VFIO to a virtual

machine.

However, the third method is not a well explored way and has some (lots of) bugs and ambiguities

(at least, for some users). But there were some people whi had succeeded at passing up to 99% of

performance to virtual machine.

Yours, Georgiy Krylov.

Comments