NVIDIA
XenServer 6.5 SP1 vs ESXi 6 - Dell R720 GRID K2
Hi It would be interesting to know if others have experienced similar results like the following and what their thoughts are... With all the "buzz" about VMware finally catching up with Citrix and supporting vGPU, I wanted to see the performance differences between the 2 latest Hypervisors as I had heard that the ESXi 6 3D experience was better than XenServer 6.5 SP1. VMware is widely recognized as the industry leading Hypervisor, but for all the extra features and (alleged) performance, you pay a substantial premium over the others, so I've always used XenServer which has done exactly what has been requested of it and performed fantastically at a price point to please all the accountants. However, it's always good to investigate other solutions, so with this in mind, I trashed my perfectly good XenServer 6.5 SP1 3D demo platform and rebuilt it with ESXi 6. The build is exactly the same as it was on XenServer, the only change in the technology stack is the Hypervisor, so after all the work of rebuilding the platform, vCenter, Gold images etc etc, imagine my surprise when XenServer out performs ESXi on every 3D demo I run. Whether the 3D desktop is Windows 7, 8.1 or 10, running it on XenServer shows higher benchmarks than ESXi. I know that these are just generic 3D apps and that you shouldn't base the entire platforms performance on some irrelevant scores and that settings have to be tailored to the specific apps that are actually going to be run, but it just seems strange that they all (so far) favour XenServer. As there is so much more cost associated with ESXi, I would have thought it should perform better than its rival. I'm running the desktops with GPU Passthrough / vDGA. At this point, I should say a little about the technology stack in-case I'm over looking something, so it's as follows: Chassis: Dell R720 (latest firmware, Drivers, BIOS etc) CPU: 2x 2.6Ghz 8 Core Memory: 256GB Storage: 2x 15k SAS RAID 1 (Hypervisor) Storage: 5x 15k SAS RAID 5 (Gold Images, Virtual Appliances) Storage: Atlantis ILIO Diskless VDI (Runs from the RAID 5 - 2vCPU / 128GB RAM) GPU: NVIDIA GRID K2 Network: 1Gb to desk (hard wired throughout, no WiFi for this test) Software: XenDesktop 7.6 FP2 Platinum (MCS to ILIO) Desktop Policies: Only base security, no other policies. I have a separate server that has all my other services on it (AD, Controller, vCenter etc etc) but this has not been rebuilt and runs on XenServer. The performance isn't hugely different visually, but when looking at the benchmarks on (for example) Unigine Valley / Heaven, Redway Turbine etc, the results always favour the desktops delivered by XenServer, and to be honest, I really didn't think this would be the case. Has anyone else experienced this? Or is there something I've missed in ESXi that's going to unleash the rest of the performance that I should have enabled or disabled? Any thoughts, comments or experiences about the 2 technologies welcomed... and this isn't a VMware bashing thread (far from it), I just haven't seen anyone do a direct comparison of 3D performance between the pair. Regards Ben
Hi

It would be interesting to know if others have experienced similar results like the following and what their thoughts are...

With all the "buzz" about VMware finally catching up with Citrix and supporting vGPU, I wanted to see the performance differences between the 2 latest Hypervisors as I had heard that the ESXi 6 3D experience was better than XenServer 6.5 SP1. VMware is widely recognized as the industry leading Hypervisor, but for all the extra features and (alleged) performance, you pay a substantial premium over the others, so I've always used XenServer which has done exactly what has been requested of it and performed fantastically at a price point to please all the accountants. However, it's always good to investigate other solutions, so with this in mind, I trashed my perfectly good XenServer 6.5 SP1 3D demo platform and rebuilt it with ESXi 6. The build is exactly the same as it was on XenServer, the only change in the technology stack is the Hypervisor, so after all the work of rebuilding the platform, vCenter, Gold images etc etc, imagine my surprise when XenServer out performs ESXi on every 3D demo I run. Whether the 3D desktop is Windows 7, 8.1 or 10, running it on XenServer shows higher benchmarks than ESXi. I know that these are just generic 3D apps and that you shouldn't base the entire platforms performance on some irrelevant scores and that settings have to be tailored to the specific apps that are actually going to be run, but it just seems strange that they all (so far) favour XenServer. As there is so much more cost associated with ESXi, I would have thought it should perform better than its rival.

I'm running the desktops with GPU Passthrough / vDGA.

At this point, I should say a little about the technology stack in-case I'm over looking something, so it's as follows:

Chassis: Dell R720 (latest firmware, Drivers, BIOS etc)
CPU: 2x 2.6Ghz 8 Core
Memory: 256GB
Storage: 2x 15k SAS RAID 1 (Hypervisor)
Storage: 5x 15k SAS RAID 5 (Gold Images, Virtual Appliances)
Storage: Atlantis ILIO Diskless VDI (Runs from the RAID 5 - 2vCPU / 128GB RAM)
GPU: NVIDIA GRID K2
Network: 1Gb to desk (hard wired throughout, no WiFi for this test)
Software: XenDesktop 7.6 FP2 Platinum (MCS to ILIO)
Desktop Policies: Only base security, no other policies.

I have a separate server that has all my other services on it (AD, Controller, vCenter etc etc) but this has not been rebuilt and runs on XenServer.

The performance isn't hugely different visually, but when looking at the benchmarks on (for example) Unigine Valley / Heaven, Redway Turbine etc, the results always favour the desktops delivered by XenServer, and to be honest, I really didn't think this would be the case.

Has anyone else experienced this? Or is there something I've missed in ESXi that's going to unleash the rest of the performance that I should have enabled or disabled?

Any thoughts, comments or experiences about the 2 technologies welcomed... and this isn't a VMware bashing thread (far from it), I just haven't seen anyone do a direct comparison of 3D performance between the pair.

Regards

Ben

#1
Posted 09/01/2015 08:44 PM   
Ben, I have no basis of comparison with VMware as we have no such installations for years, let along any equipped with GRID technology, but have seen very good performance from XenServer + XenApp + XenDesktop in our configurations. The latest generation of Tesla M6 and M60 will push the nevelope even further. We run on the Citrix platforms to a large degree because of costs. They were also the first to support the GRID GPUs and the first to support vGPU technology, and I hope they retain this edge as these products continue to evolve.
Ben,
I have no basis of comparison with VMware as we have no such installations for years, let along any equipped with GRID technology, but have seen very good performance from XenServer + XenApp + XenDesktop in our configurations. The latest generation of Tesla M6 and M60 will push the nevelope even further. We run on the Citrix platforms to a large degree because of costs. They were also the first to support the GRID GPUs and the first to support vGPU technology, and I hope they retain this edge as these products continue to evolve.

-=Tobias

#2
Posted 09/02/2015 06:45 PM   
Hi Tobias Thanks for the reply. I'm in complete agreement with you. A complete "Citrix Stack" performs extremely well with the GRID cards, and it's what I've always run previously with great results. The only reasons I'm looking at ESXi 6 (apart from just being curious about any performance differences as mentioned above) is because of hardware compatibility / support for an upcoming project we're involved with that doesn't list XenServer 6.5 on the HCL, also we have some clients that only run VMware and wouldn't be interested in a Hybrid infrastructure. Regarding the performance differences, I'm a little puzzled. It's the same hardware with the same NVIDIA (Windows) drivers, and if the results were to favour one over the other, my money would have been on ESXi to edge out XenServer due to it being more optimised. I must have missed something somewhere, or maybe Citrix really have done a blinding job! I'll keep looking at it... I'm also looking forward to getting my hands on the new 2.0 hardware, that's really going to open up a lot of different designs now that we can use blades for 3D workloads! Very excited about that :-) Regards Ben
Hi Tobias

Thanks for the reply. I'm in complete agreement with you. A complete "Citrix Stack" performs extremely well with the GRID cards, and it's what I've always run previously with great results. The only reasons I'm looking at ESXi 6 (apart from just being curious about any performance differences as mentioned above) is because of hardware compatibility / support for an upcoming project we're involved with that doesn't list XenServer 6.5 on the HCL, also we have some clients that only run VMware and wouldn't be interested in a Hybrid infrastructure.

Regarding the performance differences, I'm a little puzzled. It's the same hardware with the same NVIDIA (Windows) drivers, and if the results were to favour one over the other, my money would have been on ESXi to edge out XenServer due to it being more optimised. I must have missed something somewhere, or maybe Citrix really have done a blinding job! I'll keep looking at it...

I'm also looking forward to getting my hands on the new 2.0 hardware, that's really going to open up a lot of different designs now that we can use blades for 3D workloads! Very excited about that :-)

Regards

Ben

#3
Posted 09/02/2015 07:54 PM   
Hi, Ben: I think there have been good reasons to question this and hence, Citrix has put a lot of effort into the latest releases to address what were known issues, primarily storage and network I/O bottlenecks. Good progress has been made but I think because of the inherent hypervisor design, that VMware will still win out in some of the I/O comparisons just because of the overall design. If that really makes that much of a difference when it comes down to practical applications, is hard to say. What is important enough of a difference, really, is it 10%, 25%, 50%,...? At what point does the user even notice the difference? Is Framehawk important because of remote, not so well-connected users (limited bandwidth, packet losses, etc.) or are they all in a local lab? Can NetScaler make up some of the differences? How many of the limitations lie with the client configurations and hardware (sometimes, the last thing people consider)? There are a [u]lot[/u] of questions that need to be asked and above all, how much of a difference does cost play? If it's a Fortune 500 company, maybe not so much compared to a university struggling to make every dollar count (believe me, I know the latter situation all-too well!). Compromises will need to be made much of the time. If you have the luxury of trying out both platforms, I think that's great and it would be wonderful for the rest of us to see how things tack up against each other. The other important factor is uniformity; if something goes wrong , whom do you turn to and how good is the support? Is using a uniform environment worth it compared to trying to get the "best" pieces from here and there and coming up with a super configuration, but one that is likely to lead to a lot of finger pointing and blame when things go wrong?
Hi, Ben:

I think there have been good reasons to question this and hence, Citrix has put a lot of effort into the latest releases to address what were known issues, primarily storage and network I/O bottlenecks. Good progress has been made but I think because of the inherent hypervisor design, that VMware will still win out in some of the I/O comparisons just because of the overall design. If that really makes that much of a difference when it comes down to practical applications, is hard to say. What is important enough of a difference, really, is it 10%, 25%, 50%,...? At what point does the user even notice the difference? Is Framehawk important because of remote, not so well-connected users (limited bandwidth, packet losses, etc.) or are they all in a local lab? Can NetScaler make up some of the differences? How many of the limitations lie with the client configurations and hardware (sometimes, the last thing people consider)?

There are a lot of questions that need to be asked and above all, how much of a difference does cost play? If it's a Fortune 500 company, maybe not so much compared to a university struggling to make every dollar count (believe me, I know the latter situation all-too well!). Compromises will need to be made much of the time.

If you have the luxury of trying out both platforms, I think that's great and it would be wonderful for the rest of us to see how things tack up against each other. The other important factor is uniformity; if something goes wrong , whom do you turn to and how good is the support? Is using a uniform environment worth it compared to trying to get the "best" pieces from here and there and coming up with a super configuration, but one that is likely to lead to a lot of finger pointing and blame when things go wrong?

-=Tobias

#4
Posted 09/03/2015 04:41 AM   
We're currently working on a vGPU build based on K1 (I know it's going EOS in Q4, but I like the price). We have a big VMWare environment today, but are also running Hyper-V. We have persistent XenDesktops on Hyper-V and non-persistent on VMWare. This new build will be hosted on VMWare vSphere 6 (or XenServer 7?). I'm interested to hear, wether anyone has tested vGPU head-to-head with the latest versions of VMWare and XenServer. I'm eager to know whether XenServer still outperforms VMWare, using vGPU. Anyone that can comment on that? THanks.
We're currently working on a vGPU build based on K1 (I know it's going EOS in Q4, but I like the price). We have a big VMWare environment today, but are also running Hyper-V.
We have persistent XenDesktops on Hyper-V and non-persistent on VMWare.

This new build will be hosted on VMWare vSphere 6 (or XenServer 7?). I'm interested to hear, wether anyone has tested vGPU head-to-head with the latest versions of VMWare and XenServer. I'm eager to know whether XenServer still outperforms VMWare, using vGPU.

Anyone that can comment on that?

THanks.

#5
Posted 06/30/2016 07:53 AM   
We were asked about XenServer vs VMware on a recent webinar and you can read the reviewed answer here: http://nvidia.custhelp.com/app/answers/detail/a_id/4151/kw/xenserver%20faqs%20vgpu In short... we aren't aware of a performance differentiator but features vary e.g. GUI provision and only XenServer offers dynamic memory/overprovisioning with vGPU enabled VMs at the moment etc...
We were asked about XenServer vs VMware on a recent webinar and you can read the reviewed answer here: http://nvidia.custhelp.com/app/answers/detail/a_id/4151/kw/xenserver%20faqs%20vgpu


In short... we aren't aware of a performance differentiator but features vary e.g. GUI provision and only XenServer offers dynamic memory/overprovisioning with vGPU enabled VMs at the moment etc...

#6
Posted 06/30/2016 10:02 AM   
Scroll To Top

Add Reply