This message was deleted.

https://docs.citrix.com/en-us/tech-zone/design/design-decisions/single-server-scalability.html LoginVSI is the go to tool for load testing. With those specs roughly your looking at 8 VM's per host and 360 users per host / 45 users per VM

I would look at LoginVSI to do your load testing / baselining. One thing I would test extensively is the overcommit. We found (via LoginVSI) that overcommit KILLED our session performance. We ended up with 1:1 vCPU / pCPU ratio, minus vCPUs assigned (8) for the hypervisor (XS 7.1 LTSR). We also were doing CVAD on SBC (Server 2016) with O365, FSLogix, Chrome, WEM, PVS, on 1912 LTSR. 4 VMs / host, each 10 x 60Gb. We ended up getting about 25 users / VM, due to EMR. Your mileage may vary.

Thanks guys. @c4rm0 I’ve looked at those links before, and also the blogs that mention the rules of 5 and 10 per physical core, but alway found they never rang true. When they talk about xenApp workloads (rule of 10) are they talking published apps rather than full desktops? I’ll have a look at loginvsi aswell.

^ that rule doesn’t even consider newer CPU generations and the higher density you get

So would you say the 10x multiplyer is for published apps?

Depends on your CPU ;-) between Skylake and Cascade Lake Refresh were already seeing a density gap, let alone with older CPUs like broadwell/haswell

We’re running Cascade Lake

there is no way you will get 45 users per machine if you let them run chrome

yeah its all hypothetical as each environment's workloads will be different

Not edge?

Hi, We have HP Blade gen 10 (2 *12 pcpu + 768 GB mem), We configure CVAD on SBC (Server 2019) with O365, FSLogix, Chrome, WEM, MCS, on 1912 LTSR CU2. Create 3 VMs / host, each 12vcpu. We have about 18 - 20 users, working, everything works correctly, the processor can jump when the user is watching youtube.

Cheers for all the feedback, really appreciated

I would put together some test documentation around server builds (vCPU:pCPU ratios, pinning, hyper-threading; RAM sizing; page files / PVS or MCS caching; etc.) and then throw those against LoginVSI. Get your test workload as close as possible to your real-world workload. It will take time, and will definitely be worth it. We did that when we migrated from XA 7.15 to CVAD 1912 and found out we needed more hardware to handle the load and remain N+1 on hosts. Always better to know before hand and tell the organization, then find out after when the organization is telling you!

Just watch your MS licensing T&Cs too, overcommit can cost more $, IIRC.

2.2 GHz is slow, especialy if no GPU is being used, and especially by modern standards. We had 2.4 GHz GPUs plus NVIDIA M60 and used passthrough, one VM per fulll M60 engine, and managed 25-30 users per VM. Allocated 110 GB per VM, Windows 2016 server. Used the expensive NVIDIA Quadro (vWS) licenses, as thta seemed to make a big difference in GPU acceleration. Would now probably pick 2 x T4 GPUs.

Thanks Tobias. Out of interest what speed CPU’s are others using?

I would prefer 3.0 Ghz and up, it's always a balance between cores / clockspeed. Currently I see the Xeon Gold 6248R, 24 cores @ 3.0 Ghz at most new hosts

Agreed, @Rody Kossen. If no GPU, go for faster CPUs with sufficient cores. These days, the Intel Xeon Gold series at 3.0 GHz or faster is a good choice. And still a good choice if adding GPUs; their performance is also dictated to some extent by the CPU on the host!

What about AMD Epyc? More cores with high clock rates. Also check your BIOS settings regarding Turbo Mode and Performance settings. For Citrix Hypervisor you should select OS Control.

Epyc is architected such that those CPUs have only either 4 or 8 cores per CPU, so if using for more VCPUs than those per VM, you will see bad NUMA/vNUMA issues. See the following as well as a number of good papers by Frank Denneman: https://www.mycugc.org/blogs/tobias-kreidl/2019/04/30/a-tale-of-two-servers-part-3

Nick Rintalan updated a bunch of whitepapers on that teams site late last year https://docs.citrix.com/en-us/tech-zone/design/design-decisions/single-server-scalability.html if you dig around there's some good guides on how to use LoginVSI optimally too

@Rachel Berry, yea, VMware has/had much better tools for checking NUMA plus distributing VMs optimized for NUMA automatically.

Not just VMware 😉 #justsayin..

Yup, except I would say for old hardware.