If the devil is in the details, it stands to reason that so is his opposite number.
In the case of Microsoft’s Hyper-V, the Deiological detail is the virtual port through which data flows into and out of a virtual server to be crunched.
Because they’re not actual, physical machines they don’t have actual, physical plugs as a data input/output mechanism. Instead they have emulators—drivers that look to the VM like a physical I/O port and to the host operating system like any other data interface.
In Virtual Server—Microsoft’s current primary server virtualization product—the I/O mechanism is a driver that emulates a relatively generic Intel Ethernet card, according to Jeff Woolsey, senior program manager, virtualization for Microsoft.
“The Achilles Heel for virtualization is the I/O,” Woolsey says. “You can virtualize memory or processors because it’s relatively inexpensive and efficient. But the minute have to read a packet over wire or pull data off a disk you see the overhead and the delay. “
As an adapter it was a good choice; it supported almost everything and almost everything supported it. But it was still a straight software-emulation, an approach that “introduces a tremendous amount of overhead for I/O,” Woolsey says.
That emulation overhead came on top of the inefficiency of Virtual Server itself, which had to sit on top of another host operating system, support a guest operating system (with a virtual I/O port) which, in turn, supported the application. That’s a lot of software layers to go through every time the application has to pass a request down (waaaay down) to the microprocessor that will ultimately fulfill it.
Microsoft’s solution for Hyper-V is an entirely synthetic virtual device—an “enlightened” I/O driver, in Microsoft’s new terms.
The “enlightened driver” has a short and efficient path between data and virtual server, reducing the lag in every exchange of data by every virtual server running on a single physical host.
If there’s any single thing—other than allowing the hypervisor to run in native mode, close to the metal—it’s the efficiency of the synthetic emulated device compared to emulation of a real device.
“It’s the killer success of Hyper-V, the new architecture for I/O,” Woolsey says. “It shortens the path for networking, input and video; it allows us to really achieve a much higher level of performance.”
VMware also has I/O emulators, and benchmarks to demonstrate how effectively they work. Microsoft is offering benchmarks from partners Intel and QLogic comparing Hyper-V’s performance to Virtual Server, and showing how badly Hyper-V “smokes” Virtual Server, in Woolsey’s words.
Of course, a native-running hypervisor that didn’t smoke an emulated version would have a problem greater than the question of how enlightened it is possible for a driver to be.
Microsoft will undoubtedly put out its own benchmarks comparing Hyper-V’s performance to that of ESX, even though no one will believe them. VMware will do the same thing, with the same result. Then the two will argue about whose runs most efficiently.
But no matter how legitimate the discussion, how accurate identification of I/O as the key area of improvement, how fair or unfair the configurations for the benchmark tests, the details will matter only a little bit.
Hyper-V runs fast and efficiently, according to all the testers I’ve interviewed. So does VMware, which has a lot more users, who are a lot more than testers.
Unless one turns out to be a couple of orders of magnitude faster than the other, the benchmarks won’t matter.
The tools to manage the VMs will matter, and will continue to matter. In those, VMware is still ahead no matter how fast Hyper-V runs.
But Microsoft is well advanced in the beta process for its VM-specific management tools, and already has a solid list of physical-server management tools that can fill out its management picture. It also has a host of partners, competitors and near-strangers who have put out tools of various kinds over the years to form what Microsoft calls (not inappropriately) the Windows Server management ecosystem.
The industry has been waiting for Hyper-V to ship to see whether VMware and other virtualization vendors will be able to continue to compete against Microsoft.
The answer to that question will have to wait until Microsoft’s Virtual Machine Manager ships in a couple of months.
Until then, as with any new set of products, people will probably argue over the details and the minor differences in performance or market share or feature sets. Those don’t matter.
They’re only a distraction from the question over which set of virtualization software and associated management tools can most quickly, most efficiently and most cost-efficiently virtualize both large data-centers and mid-sized companies.
The devil may be in the details, but it’s often not the devil that causes the biggest problems. It’s people focusing too obsessively on the devil instead of on what they’re doing.
And that, in business histories looking at how well VMware dealt with this most direct threat from Microsoft, will be the detail that decides whether those histories are hagiographic celebrations of VMware’s scrap and savvy, or depressive retrospectives on the inevitable Borging of any company innovative enough to create a major change in a market in which Microsoft competes.