NVIDIA GRID vGPU on KVM hypervisors

Hello,

I have searched these forums and I found a couple inquiries in regards to utilizing NVIDIA GRID vGPU functionality under KVM-based hypervisors (RHEL/CentOS). Is this still under development? If so, is there a beta program in place to assist in testing? Any updates on progress for this capability?

I know that I can utilize passthrough for the individual GPUs on particular cards. I am looking for vGPU support under KVM hypervisors, much the same as such capabilities are available for VMware and XenServer platforms.

Thanks in advance!

vGPU support for KVM is currently under consideration.

It will require integration on their side and so I’d strongly recommend discussing this with the RHEL team’s you may work with to help them drive the business case for integration.

RedHat, Nvidia, we need this integration please!!

Hello

I saw on another thread that KVM Nvidia Grid Manager integration is a WIP.

https://gridforums.nvidia.com/default/topic/285/?comment=2476

Do you have any approx. release date ?

Thanks !

As above, it’s still under consideration and would be down to RedHat to provide any release timeline.

My comment on the other thread still stands, if and when it’s released it will be for the current release of GRID products, and unlikely to support earlier generations.

It would be really interesting to know more background on the demand for KVM and in what context you’d like to have it supported (under what stack e.g. openstack etc)… I’ll make sure this information is passed back to the product management team for them to share with RedHat etc…

So:

  • why do you want KVM vs. another hypervisor
  • what apps are you looking to use, business / office apps or £D/CAD or gaming/HPC?
  • what stack and protocols would you use … e.g. Openstack, PCoIP/RDP etc
  • what kind of use cases/environment e.g. university looking to deliver classes… service provider etc… what are you tryign to achieve and why virtualise…

Can’t wait to hear!
Rachel

@Jason, thanks for your answer.

@Rachel, you’re right, the demand may need more contextualisation. I’m working on two different project, both using KVM.

The first is a 2 weeks microscopy summer school where we use few 3D software. So,

  • KVM because it’s simple, lightweigth and foss ;
  • Specialized application, most 2D, few 3D ;
  • No stack (no broker, home-made or kvm-vdi webui). Probably Spice for remote diplay.

The second one is a larger one, we are evaluating different virtualisation infrastructure for an University. The goal is to provide server VM, on-demand services, on demand clusters (storm), and VDI for thousands of users (better if it also makes coffee). It should be multi tenant, and ACI compliant. OpenStack is one of our best choice for now.

Hope it answers your questions…
rémi

hi,

since NVIDIA have presented the Mediated Device framework on the KVM forum 2016
https://www.youtube.com/watch?v=Xs0TJU_sIPc
I take it the work to support KVM is ongoing though?
Is there any estimation available for public on when will the NVIDIA driver be available?

thank you,
/Anton

Hi Anton,

Yes it is ongoing, however such support would be dependent upon joint engineering and the timeframes of other vendors eg. ORacle/RHEL and as such we legally can’t give any estimate on timings that could influence buyign timeframes etc.

Best wishes,
Rachel

Rachel in my case I am needing to do Virtual desktops accelerated and opencl/cuda at the same time preferably on the same card. I am most comfortable with openstack/kvm/docker/lxc in order to provide the products and services I do. A foss base is much preferred to xen/vmware. For products more intended for enthusiasts as a side that is much more marketable than people having to purchase a license for home use.

I would like to add my voice to the call (plea!) for KVM GRID support. I work in research computing at a large university and we already do GPU passthrough under OpenStack for many of the same use-cases that GRID would ultimately be better for.

I’d also add that in terms of hypervisor support, I believe many deployers would be looking to offer these features in existing private or community cloud environments - OpenStack is the biggest and best choice in those spaces, and KVM is far and away the dominate hypervisor in OpenStack deployers. So I’d say there is a market of established private cloud infra that NVIDIA could sell GPUs to if KVM had GRID support.

The alternative is that AMD beats NVIDIA to the punch with their SR-IOV MxGPU. I almost hope they do, IMHO an SR-IOV approach is preferable.

Interesting … How would you add additional features to a hardware only model? Live Migration for example? …

I created and verified this account specifically to post right here and say that I really want a multi user solution for KVM. KVM has been thousands of times more stable for me than VMWare ever has. Please Nvidia. Before I’m too old to be interested in cool technology like this.

Hi Jason, is there any update about GRID vGPU on KVM? One of our clusters is based on KVM + QEMU hypervisor. We want to provide shared Nvidia GPUs for VMs, as these VMs cannot utilize a full GPU card.

Hi Xiaoning,

Currently we’re running a closed TP for vGPU on RHEL/RHV. You can expect a public release in the near future.

Regards

Simon

What does "a closed TP" mean? Test Preview?

TP = Tech Preview

I did not find released true vGPU support (not PassThrough) for RHEL KVM.

Sorry, thought it was released with latest GRID release but it depends on the latest RHEL release so it should be released very soon.