GPU Virtualization Virtualizing multiple GPUs using technology from Intel and VMware.
  1 / 2    
To whom it may concern,

I am in the process of gathering some research data for a possible system I will be presenting to my company. The goal of this system is to virtualize client desktops in our datacenter. The problem with our particular situation is that we are not just using email on the client desktops. We are in the AEC industry and our client desktop machines run heavy 3D applications that require Direct3D rendering capabilities. This makes it nearly impossible to virtualize the client machines to the datacenter.

My goal for posting in this forum is to get some more information about NVIDIA's SLI technology. I am specifically interested in the SLI Multi-OS technology that exists on the Quadro FX 3800, 4800, and 5800 GPUs (possibly more now). In September of 2010 I tried to do this but found that the only virtualization company that took advantage of NVIDIA's SLI Multi-OS technology was Parallels. The problem with the Parallels solution was that each virtual machine was tied to only 1 physical GPU. This one property killed the project for me simply because I was limited by physical hardware, and could not achieve a high VM density on each host machine.

Now, with VMware's View 4.5 and the new PCoIP protocol - I would like to know if NVIDIA's GPUs (any of them) can be virtualized by the hypervisor so that GPU resources are pooled and not mapped one-to-one, as in Parallels' solution. In addition (this is a stretch), are any of NVIDIA's GPUs equiped with a hardware encoder for Teradici's PCoIP protocol (this is to avoid the necessity of an additional PCI/e/Express slot in the host machines)?


Thanks,
-David


PS - ADMINISTRATORS: if I have posted in the wrong forum, please move this post to the correct one.
To whom it may concern,



I am in the process of gathering some research data for a possible system I will be presenting to my company. The goal of this system is to virtualize client desktops in our datacenter. The problem with our particular situation is that we are not just using email on the client desktops. We are in the AEC industry and our client desktop machines run heavy 3D applications that require Direct3D rendering capabilities. This makes it nearly impossible to virtualize the client machines to the datacenter.



My goal for posting in this forum is to get some more information about NVIDIA's SLI technology. I am specifically interested in the SLI Multi-OS technology that exists on the Quadro FX 3800, 4800, and 5800 GPUs (possibly more now). In September of 2010 I tried to do this but found that the only virtualization company that took advantage of NVIDIA's SLI Multi-OS technology was Parallels. The problem with the Parallels solution was that each virtual machine was tied to only 1 physical GPU. This one property killed the project for me simply because I was limited by physical hardware, and could not achieve a high VM density on each host machine.



Now, with VMware's View 4.5 and the new PCoIP protocol - I would like to know if NVIDIA's GPUs (any of them) can be virtualized by the hypervisor so that GPU resources are pooled and not mapped one-to-one, as in Parallels' solution. In addition (this is a stretch), are any of NVIDIA's GPUs equiped with a hardware encoder for Teradici's PCoIP protocol (this is to avoid the necessity of an additional PCI/e/Express slot in the host machines)?





Thanks,

-David





PS - ADMINISTRATORS: if I have posted in the wrong forum, please move this post to the correct one.

#1
Posted 01/26/2011 04:24 PM   
Hello there, and welcome to the forums!

What you want is unfortunately what many want and isn't really available. Parallels is currently one of the few real virtualisation solutions which supports the direct use of GPU hardware, but as you said has some significant limitations (as well as not being a very common platform).

View 4.5 and PCoIP massively improves the user experience on a VM because the actual remoting protocol is so effective, but of course you are still working off of a vGPU. Whilst VMware have long talked about implementing hardware GPU support, there is no specific timescale set on the roadmap as far as I am aware. I understand this would take significant work from both companies to achieve. Trust me, you are one of many who would like this to happen, but its not really the market VMware are after just yet.

(There is an upcoming add-in card which will offload the PCoIP compression/encryption from the CPU which will help improve density, but unfortunately won't resolve your issue.)

The lack of this capability has stopped virtualisation reaching the type of market you are in, because of the need for the power a physical graphics adapter offers, and one-to-one remoting is still really the only option.

Another option to look at is Microsoft RemoteFX which does support (and indeed requires) the use of physical graphics adapters, but still lacks some of the features required for it to be really useful for the high-end 3D market. Its also still in beta, but may be worth some investigation if you haven't looked at it already.
http://www.microsoft.com/windowsserver2008/en/us/rds-remotefx.aspx

Unlike Paralells, this actually allows you to virtualise the GPU (i.e. share its resources across multiple VMs).

"GPU Virtualization is a technology that exposes a virtual graphics device to a virtual machine. RemoteFX exposes a WDDM driver with the virtual desktop, and it allows multiple virtual desktops to share a single GPU on a Hyper-V server."

As a little background, I actually work for a manufacturer who specialises in remoting technology. Our primary market is in hardware to hardware implementations, although we have recently been doing a lot of virtualised implementations. This area is also a personal interest of mine (along with graphics of course!).

We do actually have a host card which has an NVIDIA GPU and a PCoIP chip in one PCIe slot like you mentioned, but this is intended for one-to-one and not virtualised configurations.

Happy to help out with more information if you want :)


J
Hello there, and welcome to the forums!



What you want is unfortunately what many want and isn't really available. Parallels is currently one of the few real virtualisation solutions which supports the direct use of GPU hardware, but as you said has some significant limitations (as well as not being a very common platform).



View 4.5 and PCoIP massively improves the user experience on a VM because the actual remoting protocol is so effective, but of course you are still working off of a vGPU. Whilst VMware have long talked about implementing hardware GPU support, there is no specific timescale set on the roadmap as far as I am aware. I understand this would take significant work from both companies to achieve. Trust me, you are one of many who would like this to happen, but its not really the market VMware are after just yet.



(There is an upcoming add-in card which will offload the PCoIP compression/encryption from the CPU which will help improve density, but unfortunately won't resolve your issue.)



The lack of this capability has stopped virtualisation reaching the type of market you are in, because of the need for the power a physical graphics adapter offers, and one-to-one remoting is still really the only option.



Another option to look at is Microsoft RemoteFX which does support (and indeed requires) the use of physical graphics adapters, but still lacks some of the features required for it to be really useful for the high-end 3D market. Its also still in beta, but may be worth some investigation if you haven't looked at it already.

http://www.microsoft.com/windowsserver2008/en/us/rds-remotefx.aspx



Unlike Paralells, this actually allows you to virtualise the GPU (i.e. share its resources across multiple VMs).



"GPU Virtualization is a technology that exposes a virtual graphics device to a virtual machine. RemoteFX exposes a WDDM driver with the virtual desktop, and it allows multiple virtual desktops to share a single GPU on a Hyper-V server."



As a little background, I actually work for a manufacturer who specialises in remoting technology. Our primary market is in hardware to hardware implementations, although we have recently been doing a lot of virtualised implementations. This area is also a personal interest of mine (along with graphics of course!).



We do actually have a host card which has an NVIDIA GPU and a PCoIP chip in one PCIe slot like you mentioned, but this is intended for one-to-one and not virtualised configurations.



Happy to help out with more information if you want :)





J

Official GeForce Forums Benchmarking Leaderboards

Minima: Corsair Obsidian 350D mATX, Asus Maximus VI GENE Z87, Intel Core i7-4790k @ 4.60GHz, Corsair H110, Corsair Dominator Platinum 16GB (4x4GB) @ 2400MHz, 1x Intel 730k 480GB, 1x WD Scorpio Black 750GB, 2x WD Caviar Black 1TB, 1x GTX Titan X, 1x GTX 750 Ti, Enermax 1250W Evolution, Windows 10 64bit.

Asus ROG Swift PG278Q monitor, Logitech G9X mouse, Corsair Vengeance keyboard.

Feel free to PM me if I don't follow-up on a thread that I have responded to.

Community Moderator - not employed by NVIDIA.

#2
Posted 01/28/2011 08:52 AM   
Humm, that is unfortunate. I have managed to gather a significant amount of information regarding this issue - and it looks doable, but not on VMware's platform. I think I'll have to use Microsoft's Hyper-V and RemoteFX to achieve what I need. I was able to setup a test environment last September using a Dell T7500 (fully loaded) and a NVIDIA Quadro FX 4800 with Windows Server 2008 R2, 2 Windows 7 SP2 VMs, and RemoteFX (while it was still in beta) on a thin client. The system worked nicely, but lacked the VM provisioning features I need (like PanoLogic's phenomenal PanoManager software), and RemoteFX was still just a little buggy. I'll continue my research and will put together a whole solution package - I'll post it up here when I get it so others can reference it. Right now, though, I'm looking at some pretty cool hardware from PNY for the GPU delivery to the host systems - the PNY Quadro Plex 2200 S4. I'm totally geekin' out over the PNY S4 - it's a pretty serious box!

Thank you very much for your response,
-David
Humm, that is unfortunate. I have managed to gather a significant amount of information regarding this issue - and it looks doable, but not on VMware's platform. I think I'll have to use Microsoft's Hyper-V and RemoteFX to achieve what I need. I was able to setup a test environment last September using a Dell T7500 (fully loaded) and a NVIDIA Quadro FX 4800 with Windows Server 2008 R2, 2 Windows 7 SP2 VMs, and RemoteFX (while it was still in beta) on a thin client. The system worked nicely, but lacked the VM provisioning features I need (like PanoLogic's phenomenal PanoManager software), and RemoteFX was still just a little buggy. I'll continue my research and will put together a whole solution package - I'll post it up here when I get it so others can reference it. Right now, though, I'm looking at some pretty cool hardware from PNY for the GPU delivery to the host systems - the PNY Quadro Plex 2200 S4. I'm totally geekin' out over the PNY S4 - it's a pretty serious box!



Thank you very much for your response,

-David

#3
Posted 01/28/2011 02:56 PM   
[quote name='ruready511' date='28 January 2011 - 02:56 PM' timestamp='1296226581' post='1185033']
Humm, that is unfortunate. I have managed to gather a significant amount of information regarding this issue - and it looks doable, but not on VMware's platform. I think I'll have to use Microsoft's Hyper-V and RemoteFX to achieve what I need. I was able to setup a test environment last September using a Dell T7500 (fully loaded) and a NVIDIA Quadro FX 4800 with Windows Server 2008 R2, 2 Windows 7 SP2 VMs, and RemoteFX (while it was still in beta) on a thin client. The system worked nicely, but lacked the VM provisioning features I need (like PanoLogic's phenomenal PanoManager software), and RemoteFX was still just a little buggy. I'll continue my research and will put together a whole solution package - I'll post it up here when I get it so others can reference it. Right now, though, I'm looking at some pretty cool hardware from PNY for the GPU delivery to the host systems - the PNY Quadro Plex 2200 S4. I'm totally geekin' out over the PNY S4 - it's a pretty serious box!

Thank you very much for your response,
-David
[/quote]

No worries :)

RemoteFX would seem the way to go from your description, although you should try and find out a bit more from them about their OpenGL support, which is apparently going to be a little lacking. I don't think they have publicly confirmed exactly how lacking though!

We have had a similar test environment running here, and it seems to work pretty well even in beta. The problem comes when you look at cost per position however...! The Quadro Plex system is indeed awesome, although I would also really recommend you take a look at the Dell C410x PCIe expansion rack, depending on the scale of your implementation. Recently kitted up a solution using this and its an awesome piece of kit.

If you are considering this route then take a look at the Tesla M2070Q which includes some Quadro optimisations - not sure if this is a plus or a minus for use with RemoteFX, but worth a look.

Its a shame really, cause the VMware provisioning stuff is really excellent. Haven't used PanoManager TBH, but their clients are probably somewhat under spec for what you need. What are you looking at from a client end perspective out of interest?


J
[quote name='ruready511' date='28 January 2011 - 02:56 PM' timestamp='1296226581' post='1185033']

Humm, that is unfortunate. I have managed to gather a significant amount of information regarding this issue - and it looks doable, but not on VMware's platform. I think I'll have to use Microsoft's Hyper-V and RemoteFX to achieve what I need. I was able to setup a test environment last September using a Dell T7500 (fully loaded) and a NVIDIA Quadro FX 4800 with Windows Server 2008 R2, 2 Windows 7 SP2 VMs, and RemoteFX (while it was still in beta) on a thin client. The system worked nicely, but lacked the VM provisioning features I need (like PanoLogic's phenomenal PanoManager software), and RemoteFX was still just a little buggy. I'll continue my research and will put together a whole solution package - I'll post it up here when I get it so others can reference it. Right now, though, I'm looking at some pretty cool hardware from PNY for the GPU delivery to the host systems - the PNY Quadro Plex 2200 S4. I'm totally geekin' out over the PNY S4 - it's a pretty serious box!



Thank you very much for your response,

-David





No worries :)



RemoteFX would seem the way to go from your description, although you should try and find out a bit more from them about their OpenGL support, which is apparently going to be a little lacking. I don't think they have publicly confirmed exactly how lacking though!



We have had a similar test environment running here, and it seems to work pretty well even in beta. The problem comes when you look at cost per position however...! The Quadro Plex system is indeed awesome, although I would also really recommend you take a look at the Dell C410x PCIe expansion rack, depending on the scale of your implementation. Recently kitted up a solution using this and its an awesome piece of kit.



If you are considering this route then take a look at the Tesla M2070Q which includes some Quadro optimisations - not sure if this is a plus or a minus for use with RemoteFX, but worth a look.



Its a shame really, cause the VMware provisioning stuff is really excellent. Haven't used PanoManager TBH, but their clients are probably somewhat under spec for what you need. What are you looking at from a client end perspective out of interest?





J

Official GeForce Forums Benchmarking Leaderboards

Minima: Corsair Obsidian 350D mATX, Asus Maximus VI GENE Z87, Intel Core i7-4790k @ 4.60GHz, Corsair H110, Corsair Dominator Platinum 16GB (4x4GB) @ 2400MHz, 1x Intel 730k 480GB, 1x WD Scorpio Black 750GB, 2x WD Caviar Black 1TB, 1x GTX Titan X, 1x GTX 750 Ti, Enermax 1250W Evolution, Windows 10 64bit.

Asus ROG Swift PG278Q monitor, Logitech G9X mouse, Corsair Vengeance keyboard.

Feel free to PM me if I don't follow-up on a thread that I have responded to.

Community Moderator - not employed by NVIDIA.

#4
Posted 01/28/2011 03:57 PM   
See also [url="http://community.citrix.com/display/ocb/2009/10/26/Unbeatable+solutions+for+high-end+3D+graphics"]http://community.citrix.com/display/ocb/2009/10/26/Unbeatable+solutions+for+high-end+3D+graphics[/url]
Check out this solution from Citrix for high-end 3D professional graphics virtualization: [url="http://community.citrix.com/display/ocb/2010/06/28/XenServer+Multi-GPU+Passthrough+for+HDX+3D+Pro+Graphics"]http://community.citrix.com/display/ocb/2010/06/28/XenServer+Multi-GPU+Passthrough+for+HDX+3D+Pro+Graphics[/url]
Check out this solution from Citrix for high-end 3D professional graphics virtualization: http://community.citrix.com/display/ocb/2010/06/28/XenServer+Multi-GPU+Passthrough+for+HDX+3D+Pro+Graphics

#6
Posted 01/30/2011 10:40 PM   
[quote name='Derek T' date='30 January 2011 - 10:40 PM' timestamp='1296427231' post='1186258']
Check out this solution from Citrix for high-end 3D professional graphics virtualization: [url="http://community.citrix.com/display/ocb/2010/06/28/XenServer+Multi-GPU+Passthrough+for+HDX+3D+Pro+Graphics"]http://community.citrix.com/display/ocb/2010/06/28/XenServer+Multi-GPU+Passthrough+for+HDX+3D+Pro+Graphics[/url]
[/quote]

ICA with all the bells and whistles (HDX) is certainly worth a look too, although I personally think RemoteFX may be a better solution depending on timescale.

Both are very expensive and complicated implementations of virtualisation if you are aiming for high-end 3D, especially given the type of client device required for reasonable performance. Even more so for Citrix unfortunately as RemoteFX has the advantage of being almost entirely a host-rendered solution.

This is where you reach a crucial point where you have to decide if the marginal improvements from virtualising this type of platform offer enough of an advantage over keeping a desktop or one-to-one remoting.

My take would be not just yet, as i've seen how these type of implementations go.


J
[quote name='Derek T' date='30 January 2011 - 10:40 PM' timestamp='1296427231' post='1186258']

Check out this solution from Citrix for high-end 3D professional graphics virtualization: http://community.citrix.com/display/ocb/2010/06/28/XenServer+Multi-GPU+Passthrough+for+HDX+3D+Pro+Graphics





ICA with all the bells and whistles (HDX) is certainly worth a look too, although I personally think RemoteFX may be a better solution depending on timescale.



Both are very expensive and complicated implementations of virtualisation if you are aiming for high-end 3D, especially given the type of client device required for reasonable performance. Even more so for Citrix unfortunately as RemoteFX has the advantage of being almost entirely a host-rendered solution.



This is where you reach a crucial point where you have to decide if the marginal improvements from virtualising this type of platform offer enough of an advantage over keeping a desktop or one-to-one remoting.



My take would be not just yet, as i've seen how these type of implementations go.





J

Official GeForce Forums Benchmarking Leaderboards

Minima: Corsair Obsidian 350D mATX, Asus Maximus VI GENE Z87, Intel Core i7-4790k @ 4.60GHz, Corsair H110, Corsair Dominator Platinum 16GB (4x4GB) @ 2400MHz, 1x Intel 730k 480GB, 1x WD Scorpio Black 750GB, 2x WD Caviar Black 1TB, 1x GTX Titan X, 1x GTX 750 Ti, Enermax 1250W Evolution, Windows 10 64bit.

Asus ROG Swift PG278Q monitor, Logitech G9X mouse, Corsair Vengeance keyboard.

Feel free to PM me if I don't follow-up on a thread that I have responded to.

Community Moderator - not employed by NVIDIA.

#7
Posted 01/31/2011 07:22 AM   
RemoteFX graphics hardware acceleration is limited to DirectX-based graphics; OpenGL does not go through the GPU, it is software rendered. And in the case of DirectX you should also consider the size of the 3D models, assuming you need high performance. And if you need remote access, be sure to consider bandwidth requirements, too.

The user device requirement for Citrix HDX 3D Pro is a PC with at least a 1.5 GHz CPU, or a high-end (fast) thin client, Windows or Linux. In the case of high-end 3D professional graphics, all of the graphics rendering is done on the GPU-equipped host (i.e. in the data center, not on the user device). Bandwidth recommendation is 2 Mbps or above.
RemoteFX graphics hardware acceleration is limited to DirectX-based graphics; OpenGL does not go through the GPU, it is software rendered. And in the case of DirectX you should also consider the size of the 3D models, assuming you need high performance. And if you need remote access, be sure to consider bandwidth requirements, too.



The user device requirement for Citrix HDX 3D Pro is a PC with at least a 1.5 GHz CPU, or a high-end (fast) thin client, Windows or Linux. In the case of high-end 3D professional graphics, all of the graphics rendering is done on the GPU-equipped host (i.e. in the data center, not on the user device). Bandwidth recommendation is 2 Mbps or above.

#8
Posted 01/31/2011 03:11 PM   
[quote name='Derek T' date='31 January 2011 - 03:11 PM' timestamp='1296486671' post='1186618']
RemoteFX graphics hardware acceleration is limited to DirectX-based graphics; OpenGL does not go through the GPU, it is software rendered. And in the case of DirectX you should also consider the size of the 3D models, assuming you need high performance. And if you need remote access, be sure to consider bandwidth requirements, too.

The user device requirement for Citrix HDX 3D Pro is a PC with at least a 1.5 GHz CPU, or a high-end (fast) thin client, Windows or Linux. In the case of high-end 3D professional graphics, all of the graphics rendering is done on the GPU-equipped host (i.e. in the data center, not on the user device). Bandwidth recommendation is 2 Mbps or above.
[/quote]

As RemoteFX is also not designed for the WAN, ICA is obviously going to win on bandwidth requirements.

Am I right in remembering that HDX cannot actually virtualise the GPU, i.e. share it across multiple VM's? If not then that brings us back to the OP's original issue. This will also significantly decrease the possible VM density and increase the cost per user, which I am guessing is what the OP is trying to avoid.

This high cost per position, plus the complexity of a PC (or relatively complex thin client) at the desk, and you start to lose many of the advantages that virtualisation and remoting actually offer. My personal opinion is that there isn't really a virtualisation option that can sensibly cope with this particular market segment yet, although I am sure there will be in time.


J
[quote name='Derek T' date='31 January 2011 - 03:11 PM' timestamp='1296486671' post='1186618']

RemoteFX graphics hardware acceleration is limited to DirectX-based graphics; OpenGL does not go through the GPU, it is software rendered. And in the case of DirectX you should also consider the size of the 3D models, assuming you need high performance. And if you need remote access, be sure to consider bandwidth requirements, too.



The user device requirement for Citrix HDX 3D Pro is a PC with at least a 1.5 GHz CPU, or a high-end (fast) thin client, Windows or Linux. In the case of high-end 3D professional graphics, all of the graphics rendering is done on the GPU-equipped host (i.e. in the data center, not on the user device). Bandwidth recommendation is 2 Mbps or above.





As RemoteFX is also not designed for the WAN, ICA is obviously going to win on bandwidth requirements.



Am I right in remembering that HDX cannot actually virtualise the GPU, i.e. share it across multiple VM's? If not then that brings us back to the OP's original issue. This will also significantly decrease the possible VM density and increase the cost per user, which I am guessing is what the OP is trying to avoid.



This high cost per position, plus the complexity of a PC (or relatively complex thin client) at the desk, and you start to lose many of the advantages that virtualisation and remoting actually offer. My personal opinion is that there isn't really a virtualisation option that can sensibly cope with this particular market segment yet, although I am sure there will be in time.





J

Official GeForce Forums Benchmarking Leaderboards

Minima: Corsair Obsidian 350D mATX, Asus Maximus VI GENE Z87, Intel Core i7-4790k @ 4.60GHz, Corsair H110, Corsair Dominator Platinum 16GB (4x4GB) @ 2400MHz, 1x Intel 730k 480GB, 1x WD Scorpio Black 750GB, 2x WD Caviar Black 1TB, 1x GTX Titan X, 1x GTX 750 Ti, Enermax 1250W Evolution, Windows 10 64bit.

Asus ROG Swift PG278Q monitor, Logitech G9X mouse, Corsair Vengeance keyboard.

Feel free to PM me if I don't follow-up on a thread that I have responded to.

Community Moderator - not employed by NVIDIA.

#9
Posted 01/31/2011 03:52 PM   
[quote name='ruready511' date='26 January 2011 - 05:24 PM' timestamp='1296059048' post='1183920']
I am in the process of gathering some research data for a possible system I will be presenting to my company. The goal of this system is to virtualize client desktops in our datacenter.
[/quote]

Hello,

i am also gathering iformation about the possibility to use a workstation with 4 virtual OS which should be accessed remotely. For me it will be perfect if i could assign for each OS one Nvidia Quadro 3800 Fx. Do you think it would be possible and do you have per accident probably some experience with this case? The application which i need running efficienty on the remote virtual OS will be Ansys Workbench. The work with Ansys Workbench includes rotating and moving a 3D-Geometry. Thank you for your assistance and have a nice day

Best regards
Dragomir
[quote name='ruready511' date='26 January 2011 - 05:24 PM' timestamp='1296059048' post='1183920']

I am in the process of gathering some research data for a possible system I will be presenting to my company. The goal of this system is to virtualize client desktops in our datacenter.





Hello,



i am also gathering iformation about the possibility to use a workstation with 4 virtual OS which should be accessed remotely. For me it will be perfect if i could assign for each OS one Nvidia Quadro 3800 Fx. Do you think it would be possible and do you have per accident probably some experience with this case? The application which i need running efficienty on the remote virtual OS will be Ansys Workbench. The work with Ansys Workbench includes rotating and moving a 3D-Geometry. Thank you for your assistance and have a nice day



Best regards

Dragomir

#10
Posted 02/04/2011 10:45 PM   
Another option besides XenServer Multi-GPU Passthrough (which is still in Tech Preview) would be Citrix XenApp, which runs on Windows Server 2008 R2 and supports GPU sharing for DirectX/Direct3D application acceleration. I've heard of customers running 3D apps and sharing a high-end NVIDIA graphics card between multiple users (e.g. 10 engineers).

http://community.citrix.com/display/ocb/2009/10/26/Unbeatable+solutions+for+high-end+3D+graphics
Another option besides XenServer Multi-GPU Passthrough (which is still in Tech Preview) would be Citrix XenApp, which runs on Windows Server 2008 R2 and supports GPU sharing for DirectX/Direct3D application acceleration. I've heard of customers running 3D apps and sharing a high-end NVIDIA graphics card between multiple users (e.g. 10 engineers).



http://community.citrix.com/display/ocb/2009/10/26/Unbeatable+solutions+for+high-end+3D+graphics

#11
Posted 02/12/2011 07:20 PM   
I am also looking for a VM GPU accelerate technology, currently vga passthrough has better performance, but the one-one limitation is not acceptable.
Seems GPU virtulization is really difficult.
I am wondering if opengl can declare a standard like WDDM and GPU vendor implement the standard to enable real GPU virtulization.
I am also looking for a VM GPU accelerate technology, currently vga passthrough has better performance, but the one-one limitation is not acceptable.

Seems GPU virtulization is really difficult.

I am wondering if opengl can declare a standard like WDDM and GPU vendor implement the standard to enable real GPU virtulization.

#12
Posted 04/29/2011 03:45 AM   
[quote name='Jin Cui' date='29 April 2011 - 04:45 AM' timestamp='1304048732' post='1231513']
I am also looking for a VM GPU accelerate technology, currently vga passthrough has better performance, but the one-one limitation is not acceptable.
Seems GPU virtulization is really difficult.
I am wondering if opengl can declare a standard like WDDM and GPU vendor implement the standard to enable real GPU virtulization.
[/quote]

GPU virtualisation is difficult - today there are only really three options for GPU's in conjunction with a virtual machine:

1. Stick with the vGPU provided by the hypervisor (VMware, Microsoft or Citrix)
2. Assign a single GPU to each VM (see Citrix HDX 3D or Parallels for this)
3. Share a GPU between virtual machines as a form of acceleration for performance/compression but NOT a proper 3D rendering device (Microsoft Remote FX)

When you look at these options they all have their disadvantages... with option 1 you are limited in terms of what the user can run and the performance they can expect. With option 2 you dramatically increase the cost and reduce the density per virtual machine to a point where you might as well implement one-to-one remoted computers for the performance improvements that offers. Option 3 is a step closer to the ideal but has its limitations - it also still significantly increases the cost per virtual machine, especially if you require high-end performance.

A challenging one... a standard would be nice, but I suspect this will be an initiative by one of the Hypervisor providers when it does happen. It might happen with Citrix in conjunction with Microsoft RemoteFX?


J
[quote name='Jin Cui' date='29 April 2011 - 04:45 AM' timestamp='1304048732' post='1231513']

I am also looking for a VM GPU accelerate technology, currently vga passthrough has better performance, but the one-one limitation is not acceptable.

Seems GPU virtulization is really difficult.

I am wondering if opengl can declare a standard like WDDM and GPU vendor implement the standard to enable real GPU virtulization.





GPU virtualisation is difficult - today there are only really three options for GPU's in conjunction with a virtual machine:



1. Stick with the vGPU provided by the hypervisor (VMware, Microsoft or Citrix)

2. Assign a single GPU to each VM (see Citrix HDX 3D or Parallels for this)

3. Share a GPU between virtual machines as a form of acceleration for performance/compression but NOT a proper 3D rendering device (Microsoft Remote FX)



When you look at these options they all have their disadvantages... with option 1 you are limited in terms of what the user can run and the performance they can expect. With option 2 you dramatically increase the cost and reduce the density per virtual machine to a point where you might as well implement one-to-one remoted computers for the performance improvements that offers. Option 3 is a step closer to the ideal but has its limitations - it also still significantly increases the cost per virtual machine, especially if you require high-end performance.



A challenging one... a standard would be nice, but I suspect this will be an initiative by one of the Hypervisor providers when it does happen. It might happen with Citrix in conjunction with Microsoft RemoteFX?





J

Official GeForce Forums Benchmarking Leaderboards

Minima: Corsair Obsidian 350D mATX, Asus Maximus VI GENE Z87, Intel Core i7-4790k @ 4.60GHz, Corsair H110, Corsair Dominator Platinum 16GB (4x4GB) @ 2400MHz, 1x Intel 730k 480GB, 1x WD Scorpio Black 750GB, 2x WD Caviar Black 1TB, 1x GTX Titan X, 1x GTX 750 Ti, Enermax 1250W Evolution, Windows 10 64bit.

Asus ROG Swift PG278Q monitor, Logitech G9X mouse, Corsair Vengeance keyboard.

Feel free to PM me if I don't follow-up on a thread that I have responded to.

Community Moderator - not employed by NVIDIA.

#13
Posted 05/06/2011 04:50 PM   
I'm also working on a solution which needs GPU virtualization. As a fallback option, I can accept N-N hardware VGA pass-thru tech. Could you give some suggestion for the VM choice?
I'm also working on a solution which needs GPU virtualization. As a fallback option, I can accept N-N hardware VGA pass-thru tech. Could you give some suggestion for the VM choice?

#14
Posted 05/26/2011 04:20 PM   
Ha! seems like what ur looking for is somthing they were talking about in gameinformer magazine, basically connecting to a host service (i forget what they called it) but even a low end store bought PC could subscribe to this server and play games like Crysis 2 with all settings turned up and no load would be on the client machine, and the client can enjoy 60FPS gaming, however its still in testing, and i really hope it doesnt come out, its bad enough i had to buy a Digital copy of Dirt3, i love to have my games and the cool looking box they come with :( but good luck in your search bud
Ha! seems like what ur looking for is somthing they were talking about in gameinformer magazine, basically connecting to a host service (i forget what they called it) but even a low end store bought PC could subscribe to this server and play games like Crysis 2 with all settings turned up and no load would be on the client machine, and the client can enjoy 60FPS gaming, however its still in testing, and i really hope it doesnt come out, its bad enough i had to buy a Digital copy of Dirt3, i love to have my games and the cool looking box they come with :( but good luck in your search bud

#15
Posted 05/26/2011 09:39 PM   
  1 / 2    
Scroll To Top