VDI post on Madden: good observations, different conclusions

Last month, I predicted that VDI will be just a niche play as the cloud matures. Yesterday Brian Madden posted a dramatically different perspective about the extent to which VDI will penetrate computing.

This perspective was not his own, but he thought it interesting enough to write about it. The problem though is that although the observations are reasonable, the conclusions are awful.

Let's look at specific examples.

First, the post notes that computing is changing rapidly, and of course I agree. More apps are moving toward the cloud for simplicity and portability reasons. The apps that will be left behind are rich applications that require local execution. The problem with VDI in this scenario is that you get the worst of both worlds: you get neither the simplicity of the cloud app, nor the functionality of a local app. It just doesn’t make sense to take your fat desktop and stick it into the cloud (except in niche scenarios), since VDI will only become more cumbersome as the cloud matures.

A second observation in the post invokes Moore’s law, saying that as servers become better and cheaper, the cost of VDI will drop. This might be true if users continue to use the same applications, but that’s not how computing works. Applications will continue to expand and consume the additional server bandwidth, negating any savings from Moore’s law.

Thirdly, the post goes on to describe deployment models. The primary pain point that desktop virtualization solves is desktop deployment and centralized management. With a client-based solution, IT can provision an additional VM simply by publishing an html link and sending an email. It’s cheaper, faster, and more resilient than provisioning additional boxes in the datacenter, as you do with VDI.

The post also completely ignores some fundamental issues with VDI. For example, a defining characteristic of VDI is the pooling of resources in the datacenter, but the downside to pooling is that you are magnifying the risks and complexity of desktops—the classic “eggs in one basket” problem. With VDI, you are taking inherently resilient, distributed desktops and turning them into a highly concentrated system that is vulnerable to malfunction. With VDI, if the system goes down, all your desktops go down. A related problem is that IT has to over-provision in order to prepare for peak capacity (e.g., 9:00am on Monday morning). But it’s difficult to predict group behavior, and your “over-provisioning” may prove inadequate, anyway.

Finally, the post fails to address Madden’s own Offline Paradox. Offline capability is at the core of VDI’s shortcomings. There are many times when a user finds himself without an Internet connection, such as on a plane or when the connection goes down for whatever reason. With VDI, users without a connection are unable to access their virtual desktops. This is a key area where a client-based approach excels.

What do you think the future holds? Will virtual desktops live in the datacenter or on the host machine?

Purnima Padmanabhan, VP of Products and Marketing

Back from BriForum 2010

Just got back from BriForum 2010, it was a great show as usual. This was the largest BriForum yet - both the attendee count and the exhibitor count were higher than ever. We had a table this year and got a lot of traffic. It was nice because most attendees were pretty knowledgeable about desktop virtualization and understood the benefits of client-side execution with central management, so we didn't need a lot of explaining for people to "get" the MokaFive solution. People loved our BareMetal demonstration and the fact you could manage both the BYOPC/work-from-home machines as well as BareMetal from the same management interface.

But without a doubt the best part of BriForum are the quality speakers and technical sessions. BriForum has a core of truly great presenters and speakers who talk technical and avoid FUD and marketing spin. People like Shawn Bass, Ruben Spruijt, Jeroen van de Kamp, Claudio Rodrigues, Steve Greenberg, Ron Oglesby, Tim Mangan, and Rick Dehlinger, just to name a few. And of course the man himself, Brian Madden. The presentations are great with a lot of technical meat behind them and mostly avoid the high-level fluffy marketing speak that you get at most other conferences. They are 75 minutes so you can actually get into some depth. The great presenters are what make BriForum a great event and I'm proud to have had the opportunity to present at the last two BriForums. The organizers also do a good job of treating the presenters well so I'm sure the trend will continue.

This year I did two sessions - one on BYOPC and another on Disk Workloads for Desktop VMs. The BYOPC one was in the first slot of the conference (8:45am!) and was completely full. There was a good mix of people, some of whom had deployed BYOPC, others who were interested in deploying it, and we had a good conversation. The key points were that BYOPC can reduce support costs and lead to happier users (if you do it right), and this change is happening whether you like it or not. Brian in his keynote had a great quote: "If you say there is no way you will allow it (BYOPC) in your organization, pretty soon you won't have to, because your employees will leave and go somewhere else." The other great quote I heard is: "If BYOPC is a competitive advantage today, it will be a requirement tomorrow."

The second one on Disk Workloads was much more technical. I did a deep dive into how I/O in a VM works and what a Desktop workload looks like. The desktop VM workload is quite different than server VM workloads - a typical server VM does 90% reads vs 10% writes, but a desktop is more like 60%/40% or even 50%/50%. Not only that, but the desktop VM workload is very latency-sensitive, and if you have any long latency writes, your user experience will suffer greatly. The load from a single desktop VM can peak at up to 8000 IOPS during certain operations. At the end of the session I did a demo that pitted a VM served from my Blackberry (15MB/sec read, 7MB/sec write, 10-30 IOPS) using MokaFive's optimized virtual disk format versus a normal VMDK on a much faster USB drive. The optimized one booted quickly and was very responsive, whereas the straight VMDK was sluggish, stuttering and unusable. It just goes to show that slow IO performance can make the user experience unbearable, and optimizations can make a big difference.

It was great to meet up again with the BriForum crowd and I'm looking forward to participating again next year!

John Whaley, CTO & Founder