NVIDIA Logo - Artificial Intelligence Computing Leadership from NVIDIA
GRiD vCPU on RHEL - need advice/help
(Sorry for the re-post but I couldn't find how to move or delete the original posting.) Hi All, We have the need to run two VMs (one with Windows 7, the second one with Fedora Linux 28) as KVM guests on RHEL 8 Workstation. Both VMs need accelerated graphics, i.e. OpenGL, although at very modest level. When the VMs are used, a user is sitting in front of the workstation and interacting with both VMs directly, so the local RHEL 8 desktop needs some graphics as well. My thought was to use a Nvidia card and virtualize it through GRID so that each VM sees a virtual Nvidia card it can use with regular Nvidia drivers. Having said that, I have no experience with the GRID software. So my first question would be if this can be done, and if yes, what (entry-level) Nvidia craphics card I need. Would a Quadro card suffice, or do I need something like a Tesla? Do I need two cards or can I get away with one card (which will also be used for the RHEL 8 desktop)? Thanks, B.
(Sorry for the re-post but I couldn't find how to move or delete the original posting.)

Hi All,

We have the need to run two VMs (one with Windows 7, the second one with Fedora Linux 28) as KVM guests on RHEL 8 Workstation. Both VMs need accelerated graphics, i.e. OpenGL, although at very modest level.

When the VMs are used, a user is sitting in front of the workstation and interacting with both VMs directly, so the local RHEL 8 desktop needs some graphics as well.

My thought was to use a Nvidia card and virtualize it through GRID so that each VM sees a virtual Nvidia card it can use with regular Nvidia drivers. Having said that, I have no experience with the GRID software.

So my first question would be if this can be done, and if yes, what (entry-level) Nvidia craphics card I need. Would a Quadro card suffice, or do I need something like a Tesla? Do I need two cards or can I get away with one card (which will also be used for the RHEL 8 desktop)?

Thanks,

B.

#1
Posted 06/18/2019 09:53 AM   
Hi Are the GPUs going to be located in a Rack Mount chassis or Tower chassis? ... If Rack Mount, then either a P4 or T4 would be sufficient. Rather than rewrite what NVIDIA have put together, here are links to each: Details on the P4 are here: https://images.nvidia.com/content/pdf/tesla/184457-Tesla-P4-Datasheet-NV-Final-Letter-Web.pdf Details on the T4 are here: https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/tesla-t4/t4-tensor-core-datasheet-951643.pdf Using the vGPU software, you can run both VMs on the same GPU at the same time. You will need to license each VM to use it. If you are going to use a Tower chassis, then things become slightly more tricky due to cooling requirements. You can still use the P4 / T4, however as these are fan-less, you [b]will(!)[/b] need to provide a lot of extra cooling, or you will over heat the GPU and damage it as they are not designed to work in a Tower chassis. An alternative to that would be to use a Quadro GPU (as they have a fan and are designed for Tower Chassis) in Passthrough, however as this would need to be in Passthrough, you'd then need 2 of them (1 for each VM). Basically, there are options for both, but a Rack Mount chassis gives you more options. Regards Ben
Hi

Are the GPUs going to be located in a Rack Mount chassis or Tower chassis? ...

If Rack Mount, then either a P4 or T4 would be sufficient. Rather than rewrite what NVIDIA have put together, here are links to each:

Details on the P4 are here: https://images.nvidia.com/content/pdf/tesla/184457-Tesla-P4-Datasheet-NV-Final-Letter-Web.pdf

Details on the T4 are here: https://www.nvidia.com/content/dam/en-zz/Solutions/Data-Center/tesla-t4/t4-tensor-core-datasheet-951643.pdf

Using the vGPU software, you can run both VMs on the same GPU at the same time. You will need to license each VM to use it.

If you are going to use a Tower chassis, then things become slightly more tricky due to cooling requirements. You can still use the P4 / T4, however as these are fan-less, you will(!) need to provide a lot of extra cooling, or you will over heat the GPU and damage it as they are not designed to work in a Tower chassis.

An alternative to that would be to use a Quadro GPU (as they have a fan and are designed for Tower Chassis) in Passthrough, however as this would need to be in Passthrough, you'd then need 2 of them (1 for each VM).

Basically, there are options for both, but a Rack Mount chassis gives you more options.

Regards

Ben

#2
Posted 06/18/2019 10:22 AM   
Hi Ben, First thanks for your quick and helpful reply! As to the system, it's a HP desktop workstation so the airflow is somewhat limited which I think pre-cludes any server-grade ccards that rely on massive airflow. There's also no room for any "hacked" cooling solutions, it has to be "corporate". I looked at the Tesla P4 and T4 but the cards alone are >$3k which is way outside my budget. I was thinking about using one of the lower-end Quadros (i.e. P1000/P2000) and surely there must be a way to use them for graphics virtualization on the desktop. Or am I wrong? Sorry if these are stupid questions but when it comes to GRiD software (which seems to exist in a version for server and desktop) Nvidia's websites appear somewhat confusing to me. So I'm grateful for any help/information/advice. Thanks, B.
Hi Ben,

First thanks for your quick and helpful reply!

As to the system, it's a HP desktop workstation so the airflow is somewhat limited which I think pre-cludes any server-grade ccards that rely on massive airflow.

There's also no room for any "hacked" cooling solutions, it has to be "corporate".

I looked at the Tesla P4 and T4 but the cards alone are >$3k which is way outside my budget.

I was thinking about using one of the lower-end Quadros (i.e. P1000/P2000) and surely there must be a way to use them for graphics virtualization on the desktop. Or am I wrong?

Sorry if these are stupid questions but when it comes to GRiD software (which seems to exist in a version for server and desktop) Nvidia's websites appear somewhat confusing to me. So I'm grateful for any help/information/advice.

Thanks,

B.

#3
Posted 06/18/2019 01:54 PM   
Hi As we're now talking Tower chassis, this narrows things down quite a bit and also means that you will definitely require 2 GPUs if you want to run both VMs at the same time, as none of the Quadros you'll be looking at support virtualisation. [i](Please note the term used above is "Virtualisation" and not "Passthrough". They're different ways of making the GPU available to a VM and are often referred to incorrectly, causing confusion)[/i] For reference, the only "Quadro" GPUs that support virtualisation, are the RTX 6000 & 8000, and they're both a lot more expensive than either a P4 or T4. The minimum Quadro GPUs you could potentially install and use in Passthrough would be a pair of Quadro P2000s. No NVIDIA licensing is required for this and each VM would require one of these if you wanted to use them both at the same time. Although I have not tried this myself, I have it on very good authority that the P2000 is the minimum that supports this kind of use (not P1000). Regards Ben
Hi

As we're now talking Tower chassis, this narrows things down quite a bit and also means that you will definitely require 2 GPUs if you want to run both VMs at the same time, as none of the Quadros you'll be looking at support virtualisation.

(Please note the term used above is "Virtualisation" and not "Passthrough". They're different ways of making the GPU available to a VM and are often referred to incorrectly, causing confusion)

For reference, the only "Quadro" GPUs that support virtualisation, are the RTX 6000 & 8000, and they're both a lot more expensive than either a P4 or T4.

The minimum Quadro GPUs you could potentially install and use in Passthrough would be a pair of Quadro P2000s. No NVIDIA licensing is required for this and each VM would require one of these if you wanted to use them both at the same time.

Although I have not tried this myself, I have it on very good authority that the P2000 is the minimum that supports this kind of use (not P1000).

Regards

Ben

#4
Posted 06/18/2019 02:47 PM   
Hi Ben, Thanks again for your reply. So to re-iterate: - To do GPU virtualization with Nvidia I'd need a recent Tesla card and a GRiD license - To do GPU passthrough I need at least a Quadro card per VM but I won't need additional software In a way to save some money, would this also work with older cards? For example, I could get some Tesla K20 cards for my project. Would they work, even if it means using older GRiD software? Thanks, B.
Hi Ben,

Thanks again for your reply.

So to re-iterate:

- To do GPU virtualization with Nvidia I'd need a recent Tesla card and a GRiD license

- To do GPU passthrough I need at least a Quadro card per VM but I won't need additional software

In a way to save some money, would this also work with older cards? For example, I could get some Tesla K20 cards for my project. Would they work, even if it means using older GRiD software?

Thanks,

B.

#5
Posted 06/20/2019 12:26 PM   
One other thing, for pass-through even older Quadro cards (i.e. K5000) should work, right? WOuld Geforce cards work as well? Thanks, B.
One other thing, for pass-through even older Quadro cards (i.e. K5000) should work, right?

WOuld Geforce cards work as well?

Thanks,

B.

#6
Posted 06/20/2019 12:26 PM   
Hi I don't want to direct you towards older GPUs as I'm not sure if they will work. GeForce GPUs are consumer grade, some of them used to work in Passthrough with older hypervisor versions, but I don't believe this is possible now. Whether you can find 2 that will work with your software stack depends what you have access to. Unfortunately, I don't have any experience with a K20 so wouldn't want to say yes or no to that. If you'd like to see what happens when you choose the wrong GPU, here is another forum member trying to save money by using the wrong hardware ... https://gridforums.nvidia.com/default/topic/9838/nvidia-virtual-gpu-drivers/home-server-with-an-m40-for-parsec-or-steam-in-home-streaming/ Sometimes it's just easier, faster, cheaper and more importantly less stressful, to just pay what it costs for validated technology you know will work :-) Regards Ben
Hi

I don't want to direct you towards older GPUs as I'm not sure if they will work.

GeForce GPUs are consumer grade, some of them used to work in Passthrough with older hypervisor versions, but I don't believe this is possible now. Whether you can find 2 that will work with your software stack depends what you have access to.

Unfortunately, I don't have any experience with a K20 so wouldn't want to say yes or no to that.

If you'd like to see what happens when you choose the wrong GPU, here is another forum member trying to save money by using the wrong hardware ... https://gridforums.nvidia.com/default/topic/9838/nvidia-virtual-gpu-drivers/home-server-with-an-m40-for-parsec-or-steam-in-home-streaming/

Sometimes it's just easier, faster, cheaper and more importantly less stressful, to just pay what it costs for validated technology you know will work :-)

Regards

Ben

#7
Posted 06/20/2019 03:20 PM   
-No K20 -No Geforce -Quadro >2000er seies can be used in Passthrough (check hypervisor HCL first)
-No K20
-No Geforce
-Quadro >2000er seies can be used in Passthrough (check hypervisor HCL first)

#8
Posted 06/22/2019 12:58 PM   
Scroll To Top

Add Reply