Artificial Intelligence Computing Leadership from NVIDIA
Nvidia M60/M10 Licensing VM's question
Hi Folks, We are currently working on a project to allow 6-8+ users to access a server remotely with GPU based virtual machines in an educational setting, I've looked at several solutions and am leaning towards the M60 in a DELL R740 server running Hyper-V as the hypervisor. I have limited understanding when it comes to the licensing side of this and my virtualisation experience is mostly on home labs with hyper-v and back in the day when Vsphere was just esx. So firstly I'll ask if I'm correct in the way I understand it so far. 1) Buy Server 2) Install GPU 3) Install Hyper-V 4) Install licensing server 5) Install Management software 6) Configure VM's 7) Purchase as many licenses needed for each VM 8) Fire up VM's and install software on VM's Is it as simple as that? are there any costs I've not considered? as this is an educational project in conjunction with a UK university via an epic mega grant are there any educational licensing models? The ideal setup would be 8 VM's with GPU resources that can be accessed by kids to run virtual workshops with, I have tried amazon appstream and even virtualised EC2 GPU instances but the cost is prohibitive long term. My personal thoughts are Nvidia is the way to go but I need to understand the licensing costs and options before pulling the trigger on a server and card. without going in to too much detail I have aspergers and irlens so masses of text and jumping backwards and forwards through pages is incredibly disorienting and I've tried and given up several times. So, one server, one Vgpu card, 8 virtual machines (or virtualised apps). best way to go? regards Jason
Hi Folks,

We are currently working on a project to allow 6-8+ users to access a server remotely with GPU based virtual machines in an educational setting, I've looked at several solutions and am leaning towards the M60 in a DELL R740 server running Hyper-V as the hypervisor.

I have limited understanding when it comes to the licensing side of this and my virtualisation experience is mostly on home labs with hyper-v and back in the day when Vsphere was just esx. So firstly I'll ask if I'm correct in the way I understand it so far.

1) Buy Server
2) Install GPU
3) Install Hyper-V
4) Install licensing server
5) Install Management software
6) Configure VM's
7) Purchase as many licenses needed for each VM
8) Fire up VM's and install software on VM's

Is it as simple as that? are there any costs I've not considered? as this is an educational project in conjunction with a UK university via an epic mega grant are there any educational licensing models?

The ideal setup would be 8 VM's with GPU resources that can be accessed by kids to run virtual workshops with, I have tried amazon appstream and even virtualised EC2 GPU instances but the cost is prohibitive long term. My personal thoughts are Nvidia is the way to go but I need to understand the licensing costs and options before pulling the trigger on a server and card. without going in to too much detail I have aspergers and irlens so masses of text and jumping backwards and forwards through pages is incredibly disorienting and I've tried and given up several times.

So, one server, one Vgpu card, 8 virtual machines (or virtualised apps). best way to go?


regards


Jason

#1
Posted 08/31/2020 05:15 PM   
Hi Jason If you're just getting started with vGPU my personal advice ... Don't use Hyper-V unless you have absolutely no other choice. It doesn't support vGPU, so you'll be looking at using VMs with the entire GPU passed through (DDA) to a single VM (think RDSH use case, unless you plan to give each user an entire GPU). Purely based on vGPU functionality, it is the least suitable Hypervisor for this technology at the moment. Regarding physical GPU choice as you're yet to purchase and are just getting started (assuming you choose a different Hypervisor) stay away from the Maxwell architecture (M6, M10, M60) now as the architecture is pretty old. Unless you want to use Quadro features, those GPUs will work (there's very limited Quadro support on Maxwell), but in my personal opinion it's not worth testing with them due to the overall differences they have compared to the newer architectures. The oldest architecture you should be looking at is Pascal. You could opt for a single P40, but that would be overkill, so the alternative would be 2x P4. These are 8GB GPUs and you could run 4x VMs on each GPU with 2GB of Framebuffer. However, even Pascal is starting to show it's age now in various areas, and for what you're looking to do, I'd recommend you choose something from the current Turing architecture. With that in mind, my recommendation would be a single T4 running "vPC" licensing. As the T4 is a 16GB GPU, that will give you 8x VMs each with 2GB of Framebuffer. Alternatively, if you did want to run RDSH, you could do this and use the entire GPU in a single VM with multiple students all sharing the same VM concurrently, that would also open up the option to using Hyper-V again as well. That licensing model would then change to "vApps" and you'd use the 16A profile. You would be able to support a lot more than 8x students through this model, but you're then into a RDSH vs VDI play, and that choice is up to you depending on the workload and how you want to manage it. There are reasons why you'd use one model over the other. When you buy your server, you can configure it with a GPU before ordering. This will give you a single vendor for support issues and also make sure you install the required "GPU Enablement Kit" (typically low profile CPU heat-syncs, cables etc) that are required when running GPUs. Don't forget to configure the server so that you can add additional GPUs into it later, meaning make sure you have plenty of resources installed at time of purchase as this is the cheapest way to do it rather than retro fit additional RAM or swap out CPUs etc at a later date when you want more density. Regarding discount licensing for EDU, yes this is absolutely available. Speak to your IT Partner and they'll be able to get you sorted out. If they're not an NVIDIA Partner, you can have a look here to find one: https://www.nvidia.com/en-gb/about-nvidia/partners/partner-locator/ Regarding connectivity, which Protocol were you planning to use to connect to the VMs? RDP won't work, and basically Citrix, VMware, Teradici, Mechdyne are the most commonly used and all perform very well. They all have EDU pricing options as well. Regards MG
Hi Jason

If you're just getting started with vGPU my personal advice ... Don't use Hyper-V unless you have absolutely no other choice. It doesn't support vGPU, so you'll be looking at using VMs with the entire GPU passed through (DDA) to a single VM (think RDSH use case, unless you plan to give each user an entire GPU). Purely based on vGPU functionality, it is the least suitable Hypervisor for this technology at the moment.

Regarding physical GPU choice as you're yet to purchase and are just getting started (assuming you choose a different Hypervisor) stay away from the Maxwell architecture (M6, M10, M60) now as the architecture is pretty old. Unless you want to use Quadro features, those GPUs will work (there's very limited Quadro support on Maxwell), but in my personal opinion it's not worth testing with them due to the overall differences they have compared to the newer architectures. The oldest architecture you should be looking at is Pascal. You could opt for a single P40, but that would be overkill, so the alternative would be 2x P4. These are 8GB GPUs and you could run 4x VMs on each GPU with 2GB of Framebuffer. However, even Pascal is starting to show it's age now in various areas, and for what you're looking to do, I'd recommend you choose something from the current Turing architecture. With that in mind, my recommendation would be a single T4 running "vPC" licensing. As the T4 is a 16GB GPU, that will give you 8x VMs each with 2GB of Framebuffer.

Alternatively, if you did want to run RDSH, you could do this and use the entire GPU in a single VM with multiple students all sharing the same VM concurrently, that would also open up the option to using Hyper-V again as well. That licensing model would then change to "vApps" and you'd use the 16A profile. You would be able to support a lot more than 8x students through this model, but you're then into a RDSH vs VDI play, and that choice is up to you depending on the workload and how you want to manage it. There are reasons why you'd use one model over the other.

When you buy your server, you can configure it with a GPU before ordering. This will give you a single vendor for support issues and also make sure you install the required "GPU Enablement Kit" (typically low profile CPU heat-syncs, cables etc) that are required when running GPUs. Don't forget to configure the server so that you can add additional GPUs into it later, meaning make sure you have plenty of resources installed at time of purchase as this is the cheapest way to do it rather than retro fit additional RAM or swap out CPUs etc at a later date when you want more density.

Regarding discount licensing for EDU, yes this is absolutely available. Speak to your IT Partner and they'll be able to get you sorted out. If they're not an NVIDIA Partner, you can have a look here to find one: https://www.nvidia.com/en-gb/about-nvidia/partners/partner-locator/

Regarding connectivity, which Protocol were you planning to use to connect to the VMs? RDP won't work, and basically Citrix, VMware, Teradici, Mechdyne are the most commonly used and all perform very well. They all have EDU pricing options as well.

Regards

MG

#2
Posted 09/02/2020 07:57 AM   
Scroll To Top

Add Reply