VDI Server Upgrades!
The past few weeks I've been busy deploying two new VDI severs for our Citrix XenDesktop environment to replace our existing two servers. Before I get into that, lets go back to where this journey started and how I got here...
The Beginning of my VDI journey
Back in May 2014, after being introduced to NVIDIA GRID/vGPU, I started down a journey to architect a VDI solution for our Architecture and Landscape Architecture staff that would allow us to transition from huge, expensive, physical desktops to a more flexible, mobile, and consolidated system. That journey started with a POC on XenServer with XenDesktop (it was the only one to support vGPU at the time) using a re-purposed HP DL380p G8 (being used as a VMware server host), which went through a couple different hardware iterations. Final specs below:
After nailing down the optimum CPU and GPU configuration (more on both in later posts) and executing a successful POC, I added a second server for redundancy and capacity in late 2015 along with 10 GbE switches, and a hybrid flash Dell Compellent SAN to round out the solution. Specs below:
These 2 servers have been in production for almost 2 years now, but both have fallen behind on the technology curve. Neither server are supported on newer XenServer releases, and only the Dell T630 supports newer NVIDIA Maxwell GPUs with neither supporting NVIDIA Pascal GPUs. While my original plan called for upgrading the T630 and adding a second server to compliment, the lack of support changed my approach and two new, identical servers made more sense.
Upgraded Hardware!
After waiting (not so patiently) for the convergence of Dell's latest server line release (with it the new R740) and NVIDIA's GRID 5.0 announcement which supported the new Pascal GPUs, the time to upgrade was finally here. With hardware changing so quickly, I wanted to be on the front end of the hardware refresh cycle to maximize the lifecycle of these servers. In the end I decided on two servers with the following spec:
The Beginning of my VDI journey
Back in May 2014, after being introduced to NVIDIA GRID/vGPU, I started down a journey to architect a VDI solution for our Architecture and Landscape Architecture staff that would allow us to transition from huge, expensive, physical desktops to a more flexible, mobile, and consolidated system. That journey started with a POC on XenServer with XenDesktop (it was the only one to support vGPU at the time) using a re-purposed HP DL380p G8 (being used as a VMware server host), which went through a couple different hardware iterations. Final specs below:
HP DL380p G8
2x Intel Xeon E5-2667 v2 @ 3.30 GHz (8 Core)
320 GB RAM
2x NVIDIA GRID K2 (4 GPUs on 2 boards)
2x NVIDIA GRID K2 (4 GPUs on 2 boards)
After nailing down the optimum CPU and GPU configuration (more on both in later posts) and executing a successful POC, I added a second server for redundancy and capacity in late 2015 along with 10 GbE switches, and a hybrid flash Dell Compellent SAN to round out the solution. Specs below:
Dell T630
2x Intel Xeon E5-2687W v3 @ 3.10 GHz (10 Core)
384 GB RAM
2x NVIDIA K2 (4 GPUs on 2 boards)
These 2 servers have been in production for almost 2 years now, but both have fallen behind on the technology curve. Neither server are supported on newer XenServer releases, and only the Dell T630 supports newer NVIDIA Maxwell GPUs with neither supporting NVIDIA Pascal GPUs. While my original plan called for upgrading the T630 and adding a second server to compliment, the lack of support changed my approach and two new, identical servers made more sense.
After waiting (not so patiently) for the convergence of Dell's latest server line release (with it the new R740) and NVIDIA's GRID 5.0 announcement which supported the new Pascal GPUs, the time to upgrade was finally here. With hardware changing so quickly, I wanted to be on the front end of the hardware refresh cycle to maximize the lifecycle of these servers. In the end I decided on two servers with the following spec:
Dell R740
2x Intel Xeon Gold 6154 @ 3.0 GHz (18 Core)
384 GB RAM
4x NVIDIA Tesla P4
Why the Dell R740?
In addition to being a brand we trust (less important), Dell seems to be on the leading edge of support for NVIDIA GPUs per NVIDIA's support matrix (http://www.nvidia.com/object/grid-certified-servers.html) of the big vendors. The R740's new design also allowed us to max out both CPU and still pack in 4 GPUs, which was not possible in the older R730s (hence our decision to go with the T630 before).
Why XenDesktop?
During our initial POC, only Citrix provided the vGPU compatibility. As time has gone by, they have continued to lead the charge in GPU technologies with support for Intel GPUs, NVIDIA GPUs, and committed future support for AMD GPUs (in tech preview). Also Citrix was first with NVENC (more about this in my GPU post coming up), giving the ability to utilized GPU to offload encoding from CPU.
After our Citrix POC, we did an identical POC on VMware Horizon View (after vGPU support was added). While this was also successful, user experience seemed to be better with HDX3D Pro in a variety of scenarios. At the time Blast Extreme was not an option and offloading required additional Teradici cards. Keeping an eye on both, I'm not sure either is a bad decision anymore with Blast Extreme making up a lot of ground and NVIDIA features coming to Horizon View around the same time as XenDesktop. In the end, I've chosen to focus on XenDesktop mostly due to the lack of a killer feature to make me switch and the time associated with switching and relearning everything. As a team of one (not including my helpdesk employee), I sadly don't have the time to dedicate to both.
Why XenServer?
Once we decided on XenDesktop, this was pretty easy. As a firm of 55 people, and a budget to match, XenServer was free with the licensing we already needed. As an added plus, XenServer has continued to lead the way in GPU virtualization technologies. Sometimes XenServer leaves me wanting for something more, in gets the job done and has gotten better with each release.
Coming up...
Over the next couple posts I will cover the reasoning and logic behind CPU and GPU configurations as well as comparing benchmarks relevant to our work.
Why the Dell R740?
In addition to being a brand we trust (less important), Dell seems to be on the leading edge of support for NVIDIA GPUs per NVIDIA's support matrix (http://www.nvidia.com/object/grid-certified-servers.html) of the big vendors. The R740's new design also allowed us to max out both CPU and still pack in 4 GPUs, which was not possible in the older R730s (hence our decision to go with the T630 before).
Why XenDesktop?
During our initial POC, only Citrix provided the vGPU compatibility. As time has gone by, they have continued to lead the charge in GPU technologies with support for Intel GPUs, NVIDIA GPUs, and committed future support for AMD GPUs (in tech preview). Also Citrix was first with NVENC (more about this in my GPU post coming up), giving the ability to utilized GPU to offload encoding from CPU.
After our Citrix POC, we did an identical POC on VMware Horizon View (after vGPU support was added). While this was also successful, user experience seemed to be better with HDX3D Pro in a variety of scenarios. At the time Blast Extreme was not an option and offloading required additional Teradici cards. Keeping an eye on both, I'm not sure either is a bad decision anymore with Blast Extreme making up a lot of ground and NVIDIA features coming to Horizon View around the same time as XenDesktop. In the end, I've chosen to focus on XenDesktop mostly due to the lack of a killer feature to make me switch and the time associated with switching and relearning everything. As a team of one (not including my helpdesk employee), I sadly don't have the time to dedicate to both.
Why XenServer?
Once we decided on XenDesktop, this was pretty easy. As a firm of 55 people, and a budget to match, XenServer was free with the licensing we already needed. As an added plus, XenServer has continued to lead the way in GPU virtualization technologies. Sometimes XenServer leaves me wanting for something more, in gets the job done and has gotten better with each release.
Coming up...
Over the next couple posts I will cover the reasoning and logic behind CPU and GPU configurations as well as comparing benchmarks relevant to our work.
Comments
Post a Comment