SearchServerVirtualization.com: People talk about how the hypervisor is getting commoditized, and how, in the future, the real value will lie in management software. Can you talk about that?
Bogomil Balkansky: It may be true that in the future not a lot of money will be made on the hypervisor. But when I think of the word commodity, I think of something interchangeable, that you don't care what brand you buy -- like gasoline or sugar. I don't think people will be swapping hypervisors. The hypervisor will have some customer stickiness.
We divide our management capabilities into two basic categories: Distributed virtualization is the first -- where we act on a whole farm of servers and their networks and storage. Examples of distributed virtualization products include DRS, HA, VCB and VMotion, which all act on a server resource pool.
We think there's an opportunity to provide analogous capabilities for storage and networking, and create storage and network resource pools. For example, when a VM [virtual machine] moves between ESX hosts, its [virtual machine disk format] is fixed to its storage. But what if you need to do maintenance on the storage?
This isn't storage virtualization per se; we would only build functionality as it relates to solving issues for virtual machines. Same thing with networking. We don't want to compete with Cisco, but we create virtual networks and we'll continue to go after those areas as they relate to virtual machines.
[The second category] is management and automation solutions -- for example, resource optimization and virtual machine lifecycle management, which looks at how you provision, change/modify and de-provision a virtual machine -- the kind of stuff you need to do to prevent server sprawl. We have all the core functionality today, but we want to tie in a nice end-user workflow capability.
Have you started to see people running larger workloads in virtual machines?
Balkansky: We do have some customers running one virtual machine per server today. But in general, if you're running a large workload, you don't get any benefit in terms of consolidation, even if it does help you manage downtime and availability. And if we're talking about a memory-intensive workload, in the past, we could only support 8 GB of memory per virtual machine. Now we support 16 GB. But, as we move to support 64 GB and beyond, those objections are going to come down.
I think one of the main reasons against running heavy-duty workloads [in a VM] is risk management. If I'm a customer, I'm probably going to say, this works fine for a file server, a print server.
Isn't the VMware license an impediment to running a single VM per server?
Balkansky: Depending on the application, $5,750 [list for VMware Infrastructure 3] isn't that much. If it's important enough, you'd probably pay $50,000 for the peace of mind.
What is VMware doing to improve performance?
Balkansky: There are different elements to performance. If you're talking about the latency introduced by virtualization -- anything that slows down performance is not a good thing. At the same time, [latency] is not that significant today; it's something that most end users wouldn't perceive.
Still, VMware is advancing performance by working the hardware and software ecosystem. If you look at hardware assist for virtualization, for example, Intel and AMD have a long-term roadmap for baking in hardware assist into the chips. Before, the x86 processor architecture wasn't very virtualization-friendly and VMware had to come up with roundabout ways to translate the processor instructions between the software and the hardware.
Then there's memory virtualization [Ed. note: AMD's Nested Page Tables and Intel's Extended Page Tables]. We have high hopes for memory assist. Between CPU, memory and I/O virtualization, in three years' time, we should be able to eliminate the performance overhead, so that virtual equals native.
On the OS sides, there's paravirtualization, and we will begin supporting paravirtualized Linux. One core belief at VMware has always been that you shouldn't have to do anything special, that you should be able to run any commercial off-the-shelf OS… but there will always be some segment of the Linux community that wants to do this.
Ultimately, though, stability and uptime matters more to customers than performance. If the thing runs very fast for three minutes, then crashes, then runs fast, then crashes, it's not a very good situation.
Let us know what you think about the story; email: Alex Barrett, News Director