https://worldofeuc.com logo
Join Slack
Powered by
Which machine type do people use. We are using Sta...
# microsoft-avd
r
Which machine type do people use. We are using Standard E4as v5 (4 vcpus, 32 GiB memory) but maybe there are better alternatives?
b
Good choice. You could try the new Standard_E4as_v6 - Series. Specs: AMD's fourth Generation EPYC™ 9004 processors that can achieve a boosted maximum frequency of 3.7 GHz
👍 2
j
How many users per host?
k
D4asv5 or D8 for personal D16adsv5 for multi-session - 20 users per host we typically have before running into issues with teams voice or other app issues.
j
Hm I thought Teams was AVD aware and wouldn't be an issue
n
We're using three SKUs..E16Ds (upto 20 users). D8ads (upto 8 users) and D4ads for single session depending on the workload type. Plus a ton of W365 too
E series because we're seeing RAM being a bottleneck before CPU and these have twice the RAM/Core ratio. Users having both Chrome and Edge with 1000 tabs open is a real RAM killer, even with sleeping tabs etc
j
had a demo of Nerdio and they put a proposal together.. they suggested D4_v5 (4C & 16GB) per host.. max sessions per host = 16, for 200+ users
i have a feeling im going to need more beef than that
b
@Joe Brugaletta which OS?
j
Win11 multi
👀 1
n
a D4 for 16 session can't be correct
b
16GB RAM? 1GB per user and OS is not included? no way
j
I'm all ears 🙂 knew it sounded too good to be true..
n
those numbers only work if all the users are running is notepad.exe and nothing else
😂 3
j
great 😞
right now our RDS Vms are 10vcpu w/ 64GB mem (6 of them)... on prem
Nerdio recommended many more smaller hosts vs "big" hosts, so that you could shut them down when say only 1 user is logged in at night/weekends
n
the scale out vs scale up argument largely depends on your OS+agents overhead per guest. So if you're running Defender, Tanium, CISCO AMP, Crowstrike etc in every VM, your baseline compute will be much higher, which pushes you towards fewer larger VMs. But for a baseline with a slim image with smaller oveheads, I agree, more smaller SKUs work out cheaper as you can drain them faster
how many users per VM is your on prem 10core running?
j
at peak probably 30-40 users per.. they probably need more resources at times
n
so far we've not pushed past 1.5 users per core. Performance soon drops off when you go above that we've found, although we're quite heavy with security agents which generate tons of IO when a server gets busy
j
the driving force behind AVD (well, VDI really) is to be able to run Zoom Phone/meetings (occasionally google/teams, etc) within the AVD session vs the users having to do it locally on their desktops, due to RDS's poor audio/video
n
why can't they do it locally?
j
they have to right now.
we dont allow zoom, etc to be run within the RDS sessions
n
A local experience in Teams, Zoom etc will almost always be better than AVD, even when using offloading.
j
100% agree, but at least it "works" in AVD vs its choppy as hell via RDP
I tested zoom yesterday in AVD with the VDI Client installed on the AVD host, and the VDI plugin on my desktop and it worked really well (granted.. only me on there)
so what users hate having to do, is minimize their RDS session to make a phone call, or have to copy/paste meeting link to their local pc, etc.. its sort of a separate environment
n
It does work really well these days, and Slimcore has helped alot aligning the UI and features. I use it daily.
j
we're also moving our case management system to something salesforce based, which has a zoom integration.. so if they're running salesforce in a browser in their RDS session, and the phone rings.. well.. theres no zoom so it wont pull up the contact record, etc
b
I'm also in the boat with "fewer users, more hosts". Most applications reduce functionality with >15-20 users. Mostly on newer OS.
j
it totally makes sense from "drain" ability
b
We took over @Robert Berggren's post 😂
😂 1
j
yeah sorry @Robert Berggren 🙂
👍 1
kinda related though! lol
n
back on topic..what apps are you running in AVD..just M365 stuff mostly?
j
Mostly.. outlook/excel/word/wordperfect.. browsers, pdf program (kofax/tungsten).. Zoom..
for the majority of users anyway, of course admins or programmers, etc run dif things.
n
yea I think you'll struggle to reach those density number without frustrating the hell outta your users. Modern SaaS apps are memory hungry and you'll run out of RAM pretty soon. Looking at Edge on my AVD session its almost at 1Gb for just me
j
yeah I guess i wont really know until a true POC.. but i can just see the costs going way higher
puzzled at nerdios suggestion of a D4 with 16 users though.. I would've thought they'd have a good idea 🙂
k
Reserved instances, scaling plans and aggressive disconnect and log off policies should help keep costs in check
👍 1
j
oh for sure, thats really where i felt nerdio sold itself
even things like changing the storage for a powered off host from Premium to Standard.. I wouldn't have even thought to do that (or think you could that frequently)
k
you can do all that without nerdio but nerdio is helpful if you want to use it
j
id rather not pay yet another 3rd party, but.. being as we're pretty green to the AVD idea, it seemed like a smart move.. but its not cheap
n
just check our LB policies..we now get 24 session on an E16 v5 with 16 cores and 128Gb RAM
j
wow.. way bigger vm than they recommended
how many users total neil?
n
24 per VM on that SKU
j
Right, I saw that, but was just wondering how many users/hosts you are utilizing
n
oh sorry..about 6.5k users on AVD, 3k on W365
j
awesome! 🙂 that must be fun lol.. when are you using w365 vs avd?
n
W365 for the persistent/developer use case. AVD for pooled/non-persistent.
I'd go 100% W365 if MS could make it cheap enough
j
yeah it seemed costly to go that route
hate to see that bill 😉
n
yea, cost management is fun in cloud especially trying to divvy up landing zone charges like Azure Firewall, data egress costs etc
🔥 1
k
w365 is still to expensive. We negotiated a nice deal with MS so AVD is still cheaper for multi and personal
n
if its an azure native object you can cost-attribute easily using RGs, subscriptions andTAGs but when its a shared component it's a bun fight on who picks up the bill
k
17k users for our AVD environment
j
the reserved instances is probably a game changer, but tough to commit when so early in
What about you Kyle? how many per host? what size vm? what kind of apps, etc?
if you dont mind 🙂
k
20users per vm, D16adsv5 is the sweet spot for our environment. ms apps/teams/sap and other business web apps.
I would rather add more VMs then have users submit tons of tickets for performance issues or teams voice problems so i keep it at 20users. Probably could go a little higher but its not worth it
n
It's always a trade off.,management complaining about cloud costs vs users complaining about poor performance. Pick your evil 🙂
👍 1
j
folks will have to get used to getting a message at 6pm or whatever to log out due to server being shut down for draining 🙂
i work for a law firm, so not piles of cash lol
n
you can enforce tight disconnection/logoff policies if you're doing PAYG and need to aggressively power-manage.
j
Neil/Kyle you guys do this without something like nerdio??
n
no need to message your users..just educate them around disconnection timeouts, logoff timeouts etc
we use Citrix DaaS to handle it, but its just Windows policies so can be done in Intune or GPO, or via Nerdio too of course
j
sorry I mean the VM management, not the user management
i.e. draining, etc
n
You can do it all via PoSH and graph API if you have the developer skills, but obviously Citrix/Nerdio gives you a nice console to configure it all
all depends on your staffing levels, skill sets etc. And budgets of course.
j
havent had a need for GPU VM's for things like virtual meetings, etc right?
k
Ya, all the automation is built into github actions for us and we use Service Now for a self-service portal that calls github api to trigger an action for the user like increase disk/sku/disconnect time outs/region moves/temp sku change/snapshot/restore, ect.
n
No GPUs used for most use cases, apart from a few developers working on AI stuff.
We mostly use Teams for calls and meetings, and the audio/video is all offloaded so doesn't need GPU on the hosts
A GPU on the client does help though
(as Slimcore can use local GPU for hardware acceleration)
j
and Teams just works huh.. nothing special to install. Actually, when I fired up my test AVD vm and opened Teams, it mentioned something about detecting a virtual environment and to reload
we have Teams vs business premium licensing, but rarely use it. Zoom is our phone/meeting system currently.
n
It's not quite that simple..there's a ton of connectivity requirements you need on the endpoint for it to work properly. It's all documented by MS. Not done Zoom offload but they probably also need the ability to drop code locally for the virtual channel support
j
Zoom requires a VDI client on the host, and a VDI plugin on the local pc
looking at the VM sizes you guys mentioned.. seems to be around $800/month (US) vs my D4 ~$160 lol
n
Nobody does cloud to save money...
j
aint that the truth
do you guys have local storage in those VMs for fslogix, etc?
alternatively.. dont know if horizon/citrix would be a more appropriate choice for us.. sigh 🙂
b
Depends on your usecase 😃 hybrid needed? other hyperscalers and hypervisors? TCO resides where? departmens willing to pay for their own?
l
The only real way to be sure is to test with your apps and agents, Login Enterprise can help you achieve this. Want a no obligation look at what it can do, no sales involved, DM me. I am testing D8 series in all version and found my disk activity goes crazy when running 10 users on a VM. Was able to simulate the load and see SDXHelper kicking in and absolutely trashing my performance.
n
Very smooth Mr Jeffries. But yes, ultimately everyone's environment and image will behave differently so test test test. Have you done any perf testing on ephemeral discs vs MCS ram cache?
5 Views
Open in Slack
PreviousNext
Powered by