VMware Horizon Community
justinm001
Contributor
Contributor

Is GPU basically required for horizon 7.6?

Looking to add Horizon but never was able to get it to run smooth enough for a user, and looks to be all FPS related.  We've done upgrades and optimizations to our vSAN and clusters,  also running through RDP seems to work fine.   We're mainly getting trailing windows and delay opening/changing programs, and the VMware Horizon performance tracker is shows FPS anywhere from 3-26FPS never above 26 though.  Connecting using Horizon thick client on desktop now (blast), but also issue with our pcoip 7030 and wanting to be able to use 2 1080P monitors for most users but could push to 4 1080P monitors.  All normal desktop end users so nothing GPU intensive other than web browsing, maybe occasional youtube video.  Attempted to put a few Quadro K2000 spare GPU's in server but didn't seem to help (not supported in 6.7 but no issues or errors).

Is a Grid K1 or K2 basically required for all VDI? or should I look into optimizing or other hardware related issues?

Reply
0 Kudos
12 Replies
BenFB
Virtuoso
Virtuoso

I have over 2,000 virtual desktops all connecting with 2 x 1080p monitors and not a single GPU on Horizon 7.4 and vSphere 6.0. That being said, our users may have a different expectation or need than yours.

There are a lot of factors that could be impacting you. You need to ensure your host CPU and storage are performing as well as  looking at the network. Blast and PCoIP by default both do everything in the vCPU on the guest and may need to be tuned for a ideal user experience. Once we know the hardware and environment are stable the guest OS needs to be optimized. This requires disabling unneeded applications/services and optimizing AV. We've heard that with Blast you really need hosts that have a minimum of 2.6-3.0+ GHz CPU for a smooth user experience. I have hosts with 2.3 GHz and 3.0 GHz processors with no user complaints but the user experience for power users is better on the 3.0 GHz hosts.

Reply
0 Kudos
sjesse
Leadership
Leadership

I don't have a big environment only around 60-200 users on at a time but outside of heavy video use we rarely have problems. Its a good idea to check the apps that do have issues for gpu acceleration settings, all office apps have them, and so do alot of browsers. Disabling these in pools that don't have video cards improved performance, along with what Ben says as making sure the environment is correctly sized

Reply
0 Kudos
justinm001
Contributor
Contributor

Thanks, its hard to search for info looking for not requiring GPU.  We're struggling with even a single user, but are using our cluster for other services. Everything else seems running fine so we're work on optimizing OS and testing configuration.   We're running DL580 G7 with 4x E7-8837 2.67GHz x 8 and about 512GB each, with all flash vSAN and 40GB IB NIC's, 5 servers per cluster and 3 clusters .   Its quite possible there's some other hardware issues as we have very little utilization so haven't messed with optimization yet.   We're trying off LAN and WAN too but both similar results.

We'll start working on OS optimization and testing, also build a lab with that's just hosting the VDI image.  Another question, does the view configuration server location matter? We tried both over LAN and WAN and they seem fine, just curious if we need to build another view configuration in lab or can use our one at a remote site over VPN.... I'm assuming once it connects the connection server isn't used.

Reply
0 Kudos
sjesse
Leadership
Leadership

Look at both of these

Creating an Optimized Windows Image for a VMware Horizon Virtual Desktop | VMware

Horizon 7 Enterprise Edition Reference Architecture: Validated Integration Design

Your connection servers should be in your LAN segment, for connecting externally there is an appliance to put out in your DMZ zone that those users can connect to. Thats described in the reference architecture document I mentioned. Both of these describes best practices . In general it seems they are adding alot of content at the techzone page in general for horizon that seems better then what I've seen in the past

https://techzone.vmware.com/resource/horizon

Reply
0 Kudos
BenFB
Virtuoso
Virtuoso

How are you sizing the VDI VM? The E7-8837 is an older proc with only 8 cores and no hyperthreading. With 4 x E7-8837 you only have 32 cores but you mention other server workloads. I suspect your VDI VM has high ready times which would directly impact the user experience. Are all of your VMs right sized?

Reply
0 Kudos
justinm001
Contributor
Contributor

We're trying to just get it working with a single user. Currently have a single VM with 24cores and 32GB ram with nothing else running on that hypervisor.  I've ran through some optimizations and its a lot better, but still not good enough for production, so we'll work on better optimization.  Yes the CPU's and other hardware is a bit older but still supported in 6.7, we're at low utilization so i don't belive its a CPU/Memory issue, could be a disk/network issue since we're working with vSAN and still testing with performance usage.

Reply
0 Kudos
sjesse
Leadership
Leadership

Why so big, with 8 core processors, you split across what I assume are 3 numa nodes, which will give you slower performance just there. If this is a test for one user drop it down to two cores and see how it works.All our production vms used by our users are 2 cores, and no one ever complains.

Reply
0 Kudos
justinm001
Contributor
Contributor

You're right, Trying different options to see what can help or impact performance.

Reply
0 Kudos
sjesse
Leadership
Leadership

In my experience and watching monitoring tools trying to optimize my environment, most users rarely use over a core for general use, and once you start adding more cores the cpu scheduler needs more time to work. This is because the scheduler is going to schedule all the cpus at the same time, and if there are idle cores, that vm also gets penalized. For virtual desktops is very important to right size them to just what they need, its not only cost effective, but they will generally run quicker then if they have too many cpus assigned.

Reply
0 Kudos
BenFB
Virtuoso
Virtuoso

That is never going to work unless this will be the only VM on the host. I bet esxtop will show high ready times for that VM and others. I would drop the VM down to 2 vCPU and 4 GB of memory. Than do load testing an increase as needed in small increments.

Reply
0 Kudos
schmidtl
Enthusiast
Enthusiast

Have you checked the Bios Power Management Settings of your Servers? If in doubt, try set it to "Max Power" or "High Performance" (wording is different among vendors, avoid things like 'Balanced' or OS Controlled)

The later _should_ work in vSphere 6.7, nevertheless I constantly see Environments where it is not, so before troubleshooting any further I tend to set Power Profiles to 'Max' to rule out that this is the root of my Problems.

Reply
0 Kudos
BarryUWSEFS
Enthusiast
Enthusiast

in my experience you don't need vGPU if your apps are not graphics intensive. Having said that NVIDIA now has the Tesla M10 cards which support much more dense usage that the GRID 1 and 2 with the idea of bringing vGPU to the masses. I recently deployed it with good results although the most improvement was notices with users running apps like ArcGIS and Google Earth. At any rate it might be worth a look for you.

Reply
0 Kudos