A customer asked me if we have a general rule of thumb for scoping Iops for VDI. For example, X VDI desktops require Y Iops (assuming basic MS Office use)
IOPS scaling, like bandwidth ( and to a lesser extent CPU scaling) is a very hard and exact science.
Just providing averages is deadly but designing to accommodate all the bursts is too expensive. So you will end up somewhere in the middle. That middle differs per customer. Taking averages can be risky. More experienced implementation partners will tell you about the importance of being able to accommodate bursts.
One important thing to remember is that windows xp is quite different from windows 7.
I get that, but can you at least give me some typical numbers? I've heard some tests requring as little as 3 IOPS per virtual desktop. Is that typical?
Is a good average.... in our testing I've found "work loads" to be anything between 3 and 15 for small - medium system usage. I've noticed Windows 7 booting ranges between 15 - 33 though.
I'm guessing with emerging technologies such as RemoteFX, IOP requirements will increase as more demanding applications will start hammering the disks.
Though all this depends on use case.
Other things can come into play to, disk type, storage type, raid level, speed, virtual disk alignment, etc etc.
I agree with Michel, averages are risky!
3 IOPs is generally windows twiddling it's thumbs wondering how to fill the time
Just to add... As we're using local storage for non-persistent VM's (which will count for 99.9% of our VM's), IOPs haven't been an issue for us as the servers we're using can cope with the IOPs per hosted VM. Our IOPs are spread across our hosts rather than central storage.
The headache comes when you have to make sure your SAN can cater for VDI VM's at a large scale and also make sure it's financially sustainable to scale out.
Of course this all comes down to architecture design, brick vs blade, etc...
Daniel, can you please share your server hardware configuration, i.e. make, model, CPU type & quantity, amount and type of RAM, number od local disks, type, speed, type of RAID Controller, amount of battery backed or flash cache, NIC Configuration, hypervisor type and version. I like to have cuistomer proven setups as reference to validate what we see in the lab.
IBM System x3650 M2 & M3 R2 PCI-Express (2x8) Riser Card
IBM System x3650 M2 & M3 R2 PCI-Express (1x16) Riser Card 2
IBM 675W Redundant AC Power Supply
----------------
We're using Hyper-V R2 as the hypervisor which is installed on the 2 10k disks.
The VM's are on the 12 15K drives using Raid 10
We're getting between 40 - 55 Windows 7 VM's and 1 RDSh server which is catering for 20 - 30 users (and I think it can do more).
In our rough estimations (on a flip chart), hosting 30 - 35 VM's and the RDS VM with 10 - 15 users pretty much brakes even financially. Above these numbers, we see a ROI. Adding SAN to the mix is another story though....
We plan to add FX GFX cards to each server for RemoteFX (when it comes out) so that pushes the ROI down a little but it's still there.
Bare in mind we are public sector and do get good hardware/software discounts!
Michel Roth
173 Posts
620
0
Posted September 3rd, 2010 13:00
IOPS scaling, like bandwidth ( and to a lesser extent CPU scaling) is a very hard and exact science.
Just providing averages is deadly but designing to accommodate all the bursts is too expensive. So you will end up somewhere in the middle. That middle differs per customer. Taking averages can be risky. More experienced implementation partners will tell you about the importance of being able to accommodate bursts.
One important thing to remember is that windows xp is quite different from windows 7.