VDI and SAN / NAS performance

[myxiplx]

Is there anybody out there running VDI who can provide some figures as to how it affects the SAN / NAS?

We're looking at a couple of alternative technologies to provide a cheap NFS based NAS for VDI, but I'm struggling to size the new server without resorting to guesswork. Does anybody have some figures of what's needed in terms of raw bandwidth and IO's per second for running Windows XP workstations? I appreciate everybody will get slightly different mileage depending on usage, I'd just like a rough idea of what kind of network & disk performance I'm likely to need for 100 workstations running off a single NAS.

thanks,

Ross

[williambishop]

Depends on the workload of the XP desktop. A standard office or clinical system, the IO is not very heavy at all. I rarely ever see it average over 5 IO's and 70k in bandwidth, during installation of a new application, it will spike to around 30 or 40 meg a second, but that is not a common thing. Now, I do have some machines that do 200x that, and I have some that are even lower that listed. I personally was shocked at how minimal the impact was on the san side. I have over 2000 xp desktops and their cumulative is less than I see on one mainframe.

Now, mind you, you will see mostly random reads and writes. We're not pulling tracks for an application server after all.

[myxiplx]

That's great, good to see some actual figures. Would you mind sharing the total load on your SAN from your 2000 desktops?

[myxiplx]

Oh dear god. I just ran some figures based off that. If our workstations loads are as low as yours, we can take the worst case scenario from our theoretical NAS performance figures and our network storage is still only setting us back £30 per desktop. If we look at the optimistic figures it's as low as £10 per machine.

Ok, we're not on redundant storage yet (we're still trialing VDI), but I suspect we're looking at fully redundant storage figures of around £50 a machine.

[williambishop]

I'll pull some xp desktops disk load from vmware. It's difficult to pull from the array, as there are literally hundreds of connections to it, of every flavor(aix,windows,linux,esx, etc) and purpose(servers,desktops).

I'll pull about 5 desktops as graphs, give you a 24 hour snapshot and send them as pdf. Will that work for you?

[williambishop]

That might be wasteful...For my servers, I have redundant storage, so I can move them if necessary. But I have no plans to do the same with desktops, which are quite literally disposable. My plan is to keep 15% of them on another array, enough to keep the systems going in case of an array failure, but desktops aren't as critical, and I can spin off new ones pretty rapidly(about 15 minutes each).

If you can keep your image size down as well, you'll drive that price much lower. I keep mine at around 5gig, and after a dozen apps, it's still around 1.5 gig free(enough for a couple more apps). But then again, I use XP which is an easier load than others who have used vista from what I hear. I do get 150+ desktops per terrabyte.

[myxiplx]

That PDF would be superb, and very much appreciated.

I can see what you're saying about redundant storage, but we're very concious here that we're moving from a system where a component failure affects a single user, to one where a failure affects dozens. We probably among the smallest companies to be rolling out VDI, we don't even have a SAN for our servers yet, so this is very new ground for us.

However, it's looking like our NFS storage server for VMware isn't going to be expensive, and my thinking is that I'd rather have two relatively cheap NAS servers, than one expensive one. Right now I'm looking at £2,000 for a 2.5TB server that should easily cope with 100 clients. That's our temporary solution while we see if VDI is going to be suitable, but I don't envisage spending much more on the final servers. If the right hardware support comes along over the next 12 months we'll be buying a pair of servers for around £3,500 each which should be fast enough to cope with all our storage demands. We're actually planning to give the end users a more responsive desktop by switching to VDI.

Talking about image sizes, have you seen the news about VMware's Scalable Virtual Image technology? One image for many desktops: . I'm hoping that might reduce things enough that we could consider flash storage instead of fixed disks when we come to buy the main servers.

Indicentally, what do you do with the windows swapfile? Is that included in the 5gig image size you have there?

[williambishop]

Here's a quick 3, just pulled at random. As you can see, while there are spikes, the average is VERY low. These are on nursing stations, manned 24x7, and since we don't use physical charts, these are used for everything from charting, to writing documents.

BTW, those are all 24 hour graphs.

[mittim12]

Do you use any other methods outside of Virtual Center to monitor IO?

[williambishop]

Added, one desktop only LUN, IO's for 24 hours from performance manager.

[williambishop]

Typically, I use control center and performance manager, but it's too big an ordeal to pull all the desktops from it, since the desktops aren't the majority of what I have on it...I pulled one clusters' lun for post right above this.

[williambishop]

I neglected to address some of the other points(I had maitenance this morning to do)....So here goes.

It can be a more responsive desktop. I don't know about using NFS or iSCSI, latency would be a bit higher than FC attached, but not much I'm sure. Ours are pretty responsive, even more so those that are attached with the wyse v10L's.

Image size from scalable virtual image. I can't discuss anything under NDA.

Swapfile, I always store with the vm, it's easier to manage. yes it's included. Surprising yes, but once you find that you really need a lot less memory(we use 296 where on physicals we used 1-2 gigs), you'll find your swap needs also decrease.

Flash storage is still evolving, and it is still prohibitively expensive, especially the enterprise stuff like EMC uses that has a MTBF similar to spinning media.

[myxiplx]

We're going to be testing some 10Gbps Infiniband cards here, and we're hoping that even with NFS latency will be pretty low. We've got 512MB of battery backed nvram for the write cache on the server, and 8GB of read cache. We're hoping that between all that it will make for a very responsive client, but we won't know for sure until the hardware arrives and we start testing later this month.

I guess you've already heard about the scalable stuff, and I agree, flash is expensive now, but I can see prices tumbling rapidly, and if VMware can reduce the footprint needed for VDI, that will also have a massive effect on the cost of the flash memory you need. Fusion-io have an 80GB PCI-e drive for around $2,400. With 5GB images and allowing 1GB per machine for differences you could still fit 75 vm's on that. Throw that in a cheap server with a $125 Infiniband card and I bet even with NFS your clients would fly.

[williambishop]

I'm thinking about a year before flash drops enough to be usable.

I'm expecting about 80% reduction in space with the vmware product upgrade, but IO's will go much higher with this setup.

[PghMike]

Sorry for the sorta-unrelated question, but can you direct me to the source of the 512MB of non-volatile memory?

Thanks,

Mike