Tesla V100 16 GB currently on Hyper-V with DDA - Any advantages using Quadro vDWS?

Hello.
I’m new to all this vGPU world and I received a server with 4 Tesla V100 (16GB). I had to install windows server 2019 with Hyper-V and configured it with Discrete Device Assignment (DDA) because, from what I read until now, there is yet no way for hyper-v to distribute the GPU across all VMs. But, I also read that a microsoft solution might be underway soon.

Today I came across Nvidia vGPU. Supposedly, Quadro vDWS should be the software to install (the server hardware is fully supported), but does it still require Hyper-V with DDA? If so, does this mean I can only have 4 VMs using the 4 GPUs available? If that’s the case, I don’t understand the purpose of using Quadro vDWS. My goal is is to make the 4 GPUs available to all existing VMs.
Thanks and regards
David

Hi

The problem is Hyper-V, not Windows, NVIDIA, the V100s or vDWS.

If you switch Hypervisor to one that supports vGPU, you’ll have far more functionality available. XenServer, vSphere, KVM, AHV, they all work with vGPU. Hyper-V is the one that doesn’t. You’ll need to use DDA (Passthrough) if you want to carry on down that road, and yes, that will limit you to 4 VMs, 1 per GPU, as you can’t virtualise them (with Hyper-V).

As for vDWS, that’s part of a 4 tier licensing model. vApps, vPC, vCS, vDWS. Each version allows different functionality, with vDWS being at the top.

Out of interest, what was your planned use case?

Regards

MG

1 Like

Thanks for the reply. So if I understand it correctly, even if we buy the vDWS package it won’t be possible to make the 4 GPUs available for all VMs? But if so, I can’t see any advantage of using it with Hyper-V and still it is available for purchase?!

On the "NVIDIA VIRTUAL GPU PACKAGING, PRICING AND LICENSING" document I read yesterday (from June 2019) Hyper-V is supported and the recommend license is Quadro vDWS.

Although I was not consulted for que acquisition, this server is supposed to be used at a University for department for research.
Thanks and regards.
David

Hi

At the moment, the only option with Hyper-V is to allocate an entire GPU to 1 VM, or multiple GPUs to 1 VM. As there is no vGPU, there is no ability to run multiple VMs on a single GPU.

Whether used in DDA (Passthrough) or vGPU, Tesla GPUs require a license depending on the use case and required functionality. With Graphical sessions, the licensing is per concurrent user. With Virtua lCompute Server, the licensing is per GPU. Have a look here for license functionality: Client Licensing User Guide :: NVIDIA Virtual GPU Software Documentation

As you’re using Hyper-V, your options are limited to either allocating 1 V100 to 1 VM and having a total of 4 active VMs with 1 user on each. Multiple V100s to a single VM, or, you can use a VM with the RDSH role installed and a V100 attached and provide multi-user access like that to achieve greater density. Changing the Hypervisor will give you vGPU functionality.

How many VMs do you have that require GPU access?

Regards

MG

1 Like

Thank you so much for the reply. For now we are using 3 VMs, but I guess it won’t be long to all 4 have their respective assigned GPU.

In the link you provided

Table 2. NVIDIA vGPU Software License Enforcement by Deployment Type

Microsoft DDA for workstation or professional 3D graphics Quadro vDWS Software
Microsoft DDA for compute-intensive virtual compute servers Software See Note (4).
Microsoft DDA for PC-level applications GRID Virtual Applications EULA only

but in this document https://images.nvidia.com/content/grid/pdf/Virtual-GPU-Packaging-and-Licensing-Guide.pdf (updated since yesterday with Red Hat KVM :) )

There is this section.
I am using… I need this license…

Microsoft
Hyper-V (DDA) Quadro vDWS

Previously in the same document VGPU software editions and entitlement also show Quadro vDWS as a more complete solution, am I right? Anyway, no one considered any licensing, so I guess users will have to use Hyper-V with DDA and none of the vDWS or Virtual Computer Server. GRID Virtual Applications seems the only one to be available for free.
If I got it right anyway…

Hi

Yes, QvDWS is the premium level license. This license covers everything the GPUs are capable of delivering, from performance to features and functionality. There is no higher level license.

DDA / Passthrough will still require a license. Have a look at the table on Page 12 of your link above for guidance on specific use cases. You can either use vApps or QvDWS depending on your requirements. QvDWS will allow CUDA workloads and Quadro features, whereas vApps is for RDSH and more basic applications.

Most expensive scenario, assuming 1 VM per GPU, you’ll need 4 QvDWS licenses, but we’re not talking about a lot of money in the big scheme of things. That said, in your post above you mention the workloads will be run at a University, therefore I assume you work in EDU, if that’s true, then you could be entitled to an EDU discount for any licenses, so check with your license provider before purchasing :-)

Regards

MG

Hi.
Thanks again. yes I read about education licences.
The weird thing is that I have already at 2 people working with 2 linux VMs using the assigned GPU, and everything seems to be working fine. No one told me otherwise. I guess they might be using some other toolboxes/driver/open source software that doesn’t require any of these softwares?
Probably I can only get some more information during next week, as some of them are still on "vacations".
Thanks and regards.
David