https://worldofeuc.com logo
Join Slack
Powered by
What kind of user density are people seeing in the...
# citrix-vad
r
What kind of user density are people seeing in the field for Server 2022 and CVAD 2402.... I'm seeing like 8 users on a 4/24 system. CPU is just getting hammered. CPU is only 4 years old.... - Full desktops.
j
What's doing the hammering?
r
normal shit
chrome/edge/explorer
d
GPUs? If not. Disable hardware acceleration 👍
r
yup, standard GPO in our stuff - Though, i will trust but verify. 😄
To be clear.. WHICH place do you mean?
There's office/chrome/edge... I think you can do it at a Windows level?
d
Havent heard anything specific to 2402/2022. Are you using ad-blockers?
Oh sorry. For each product. Edge, Chrome, whatever
r
thank you, sir
j
Also you could enable browser content redirection for certain sites to offload that stuff to the endpoint: https://docs.citrix.com/en-us/citrix-virtual-apps-desktops/multimedia/browser-content-redirection.html Won’t do anything for explorer though.
r
yeah, we BCR things as well. I have (what i feel like) are all the community best practices... but, since going from 2008R2 to 2022, user density has plummeted.
What kind of user density are you guys seeing?
j
Newer OS generally ends up with fewer users per server. From what yours saying 8 users on a 4vcpu machine is still pretty good, would you say their light workloads? We run 8vcpu and get around 10 users (server 2016), 2019 pretty similar, but they are pretty heavy, with lots of dodgy apps. What hosts are you running, how many pCores, and what speed are they (normal/turbo)? We run leveno, 2.2GHz/3.4GHz.
r
multiple different PODs... some Xeon 5218s, some Xeon 6346. Newer hardware has EPYC 7543, which certainly perform better. But, sounds like you have the same density that I have, basically.
j
I’m pretty sure our workloads are well optimised, but would be interesting to hear what others are seeing for their deployments?
r
I remember 30+ users on Server 2000/2003. 😄
j
So do half of our clients 😂
this 1
m
We were just discussing that today. We took a big hit going from 2012r2 to 2019 and I told them we will have less performance going to 2022 as well. It has never gotten better going to a new OS.
r
yeah, but 6-9 users per server.. OUCH
👍 1
j
All depends on the workload, light, medium, heavy. But with Flashy Office apps, Teams, dodgy browsers etc, I think light and medium workloads are as rare as rocking horse sh*t when doing full desktops. I think vendors have forgot about desktop virtualisation and just assume their apps and OS are being deployed on physical, multi core, GPU enabled devices with loads of RAM
j
Downgrade for density, makes sense
🤔 1
j
Think I’ll float the downgrade option. They’ll think I’m mad until I point out we can get tripple the users per server 😂
d
2019 servers and 8 is like max. i think 5 people per server is sweet spot which is dumb. Maybe we need to invest into GPUs
r
I've been moving clients to published apps, because it's just crazy the scale
d
GPUs is almost a necessity nowadays! Especially consideringer that the Office Suite is produced for the consumer, who always has a GPU available.
j
GPUs for Office workloads. Ludicrous, yet here we are
j
Yeah, and you don’t even get an animated paper clip anymore!
d
Rankin for pres 😎
l
what proc?
d
was hoping the one benefit to new teams would increase user density but haven't really seen that. our proc is - Intel(R) Xeon(R) Gold 6346 CPU @ 3.10GHz
l
new teams with offloading is MUCH better than classic teams as far as density goes from what I've seen. That's a good clock speed proc, but VERY low on cores
In general, from what I've seen, Cores relates to density, clock speed to overall performance
d
yea we needed as much speed as possible due to Bloomberg and Outlook. if speed is lower than 3GHz then Outlook lags like crazy
l
In general though, I would say 8 users on a 4 core VM is the best youre gonna get. How many VMs per host? Doing good if you're getting 4 per host I think
d
thats a touchy subject due to disagreements in how to provision cores. at the moment its at an absurd 12 core per VM
part of me feels like there is a performance drop from cpu wait times but alas
l
buddy was doing testing in azure and on an 8x32, it was getting hammered on login with 5 people
Is your env @Danny different than Ryan's? I saw he said a 4/24 system
12 cores per VM seems like a lot if its idle, but with a ton of users, I could see it
d
ranges from 5-8 per server
sorry for my ignronance but dont know what 4x24 or 8x32 mean. 8 cores and 32GB ram?
✔️ 1
l
@Daniel Madsen doesn't chrome disable hardware accel automatically when it detects a server os?
d
12x128 in that case
l
12x128 and only 5 - 8 users? oof, that seems really low
but how many VMs per host?
d
we try to keep each around 4-6 per host. sometimes need to rotate around based of host CPU usage
j
I try to stay under half NUMA and 2:1 overcommit (pCores) max, so on 36core (2x18) would take 4 pCores for hypervisor overhead. Then with the remaining 32 deploy 8 x 8vCpu session hosts. This seems to work ok, although to get a solid number then your best using something like login vsi to validate your configuration
j
I have 2 * EPYC 7x something per box no CPU overcommit, bad idea for RDSH, 8vCPUs per VM, 8 users per VM, 16 VMs per host. We are on W16, heavy old apps. We are moving to W22, I hope the hit won't be too bad ...
j
@jvitech how many physical cores do your hosts have?
d
@langsbr is that new? Never heard that.
j
@John bullough I have 2 physical cores per box, 64 vCPUs per physical core with Hyper Threading. So total per box 128 vCPUs. 16VMs per 8vCPUs = 128.
j
For multi session I START at 8 vcpu x 32 gb of ram. And go up to 16 vcpu x 64 gb of ram. But every environment is different, so testing is necessary. With 8 x 32 you can get 50 users running notepad, but maybe 4 running huge excel BI data set manipulation. Also look into using WEM for "sleeping" disconnected sessions, and enable browser tab sleeping for better resource return.
👍 2
d
for some reason i felt that 16 made things worse. maybe i should try that.
r
The NUMA stuff has changed so much in the recenter Hypervisor versions, i have no idea anymore
d
Which hypervisor, any other workloads on the same hosts? We usually configure 8vCPU with at least 32 GB RAM. The newer Epyc machines with 32 pCPU (64 HT) do a great job. We often use XenServer with NVIDIA A16 GPU‘s. So, with a 8 GB framebuffer profile you get exactly 8 VDA on one host without doing so much overprovisioning. We get about 10-14 users on each VDA. If you use XenServer there is a great way to monitor CPU usage and make sure the machines has the right amount of vCPU‘s. https://docs.xenserver.com/en-us/xenserver/8/monitor-performance/cpu.html
r
Hyper-V. Yes, generic workloads on all clusters.
m
Can confirm that our density is also very low. Average customer has 4 vCPU and 38 GB RAM with about 8-9 users.
r
Need a hard push to published apps, me thinks
l
The last place I was at started on physical PCs, moved to XA published desktops on bare metal, then to VDI and kept wondering why costs kept increasing. I tried to move to published apps (99% of everything was web based). No traction...
r
these density numbers are so expensive..
l
yup - it was expensive even with a baller price from HP and Nutanix. somehow the thought is azure will be cheaper? lol
r
LMFAO... wtf thinks that?
j
Funny enough CIO know it's more expensive but still want to go Cloud, they feel like they have to, mainly to take advantage of cool products, AI being one of them. Speaking of AI, when do you think we will be replaced ?
j
AI is not just in the cloud, surely?
m
I hate the word AI. It’s all LLMs or machine learning? Haven’t seen any AI yet.
j
Yep ML is the right term
l
funny part is - previous mgmt did the whole "public cloud is the way" into AWS - it was WAY too expensive, went back on-prem. New Mgmt comes in with brilliant "Azure is better, trust me yall" idea. They just haven't gotten the bill for 75K concurrent users yet
j
First part of your cloud strategy should be your exit strategy (I believe @Kees Baggerman coined this phrase, or similar)
❤️ 4
j
Machine learning is not the same as Artificial Intelligence. I was thinking about an AI that could do our job, we will come to that especially when all be Cloud and normalised. I am not pro-Cloud, quite the opposite. AI on-prems can be done but that's not that easy, need load of compute ... but yeah they try to find any excuses to go Cloud. I don't know what's their agenda but they are all so pushy with Cloud 🤔
r
is compliance that's going to drive Cloud. When all us regular datacenter folks can't afford to put in the requirements.
then Cloud will just SUPER jack up the prices... I hope to be retired by then
2 Views
Open in Slack
PreviousNext
Powered by