This message was deleted.

I'm lucky to see 10 users on a 4/24 system.

Yea same here. On 2016 we get 17 users on server .with the same published app across 20 VDA. 4vcpu and 20gb or Ram. Pvs Write cage is 3Gb. Published application is very light weight though. 2019 that is used for published apps and published desktops 6vcpu and 24gb of ram . pVS write cache here is 4096 we get on an average 5-9 users per servers.

It doesn’t seem to matter how high you spec it, if you’re doing full published desktops and letting people run a browser then it craps out at about 15-20 sessions.

Mikes right though, apps are hogs now. You give it to much cpu then cpu wait times occur on the hypervisor. Then you get into vNuma/Numa miss matches and causes cpu response times to be higher.

what you tend to see on these shared servers though is that just about all the load ends up on cores 1 and 2

Awesome feedback, this aligns with what I've seen and have been dealing with.. oversized RDSH Host are no good and is best to scale out vs scale up. Large VM causing High ready times and co-stop on the hypervisor

I personally think WEM can help on some of this as well .

I need to look into WEM CPU stuff mroe

WEM helps me cpu spike/ clamping and the new stuff to release resources back on disconnect sessions. There are a many T shits size now a days to this madness lol.

Shared server desktops are not fun if those users are leveraging apps that hog CPU. It's a reverse Oprah effect. Everyone else gets to suffer.

That's very true

99.9% of the time it's browsing.

Browsers suck

i really try to abide by the rule of never over allocating the virtual (hyperthreaded) core count, for Server based workloads.

Don't forget MS teams when you're talking about browsers, not surprising with it being electron. Top three resource "consumers" are always Chrome, Edge, MS Teams.

12 users on a 6 core / 32 GB ram but thinking to limit it to 10 users max now. And probably increase the ram to 40 GB.

I feel like Meltdown/Spectre and all the side channel stuff in combination with Server 2016 and subsequently Server 2019 has just made it so the same VM "size" goes about half as far as it used to. Getting our first gear with 3rd gen Xeon Scalable soon, I'll be interested how VM sizing goes...

💯

Do we talk about RDSH? We have very good results with 2x16 Core with hyperthreading enabled. That gives you 8 VM‘s with 8 vCPU and at least 32-40 GB RAM and 6-8 GB PVS Cache to RAM. In most cases we get 18-22 sessions on a VDA. So, ~ 150 sessions per host. With the right settings you can handle Edge (sleeping tabs, uBlock, etc.). WEM helps with CPU management. 4 x NVidia T4 fits perfectly (8 GB framebuffer per VDA).

Does 8cpu help any more than 4 or 6 would?

Yes, definitely! The point is that you don't overcommit your physical CPU’s

To the point where controlup told me to scale back to 3 cores

Have seen this too, but in this case I trust the users.

hmm. not trying to derail but ControlUp insights is telling us our win10 VMs are oversized yet the realltime monitor tells us a completely different story. But then insights tells us we have double the number of hosts we actually have also.......

Yea CU told me as well not to do 8 and we backed it to 6. But that was based on the Numa parts in the CPU/Ram Architecture.

Regarding NUMA the 2x16 core 8 vCPU is really a good way to go.

I tried so many different things in terms of cores and ram but it’s going to be different from place to place. I went with 6 because it aligned better with the core count on our processors

All good points and lessons from the field. I appreciate it. I've been running an environment where we still looking for the sweet spot. We seem to definitely hit a brick wall at around 25 sessions. We did a 1.6 pCPU/vCPU core ratio but the VDA were 16 vCPU 102vRAM (2x20 Cores pCPUs 512GB pRAM ESXi with HT enabled) and the ready times and co-stop at peak time were high. Then also dealing with other software issues etc definitely not making for a good experience. Reduced to 12vCPU/77GB vRAM and we are still getting about the same number users and experience seems to fair better but not great yet. But sounds like we may be cramming too much. Especially since most of our users are knowledge to power users.

8/32 MCS on W2016, max 15 users. Heavy browsing environment

we are 12x96 and get up to 50 users and what seems like still some head room. Mostly Chrome users PVS with 2GB RAM WC WEM with CPU mgmt 2016 server with latest VDA

@Mike Streetz (O_P) so you are saying any more than 4 CPU on a 2019 server is a waste?

it really depends what you are doing but for a desktop with office and web browsing we did not see enough of an improvement to make adding cores better than spinning up a whole new server

interesting. Any special agents in use, Mike?

like AV? It had crowdstrike on there but that’s pretty low utilization

CS has a shockingly low footprint. We came from Mcafee and couldn't believe our eyes.

For Virtual apps servers that are delivering heavy graphics workloads the baseline is 8 vCPU and 32GB of RAM which gives us 7-10 users/vm and the 6q profile for the NVIDIA RTX 6000. Workloads are getting more demanding so we increased the RAM to 48gb. and are looking at moving to the 8q profile because 6 isn't supporting hardware encoding on the CVAD 2203 release.Pay attention to NUMA and remember that overprovisioning vgpu can actually slow down performance.