During his talk, Dike discussed recent work to allow UML to use hardware virtualization extensions added to x86 processors by Intel and AMD. Dike says that the extensions are largely compatible between Intel and AMD CPUs, and will allow UML to intercept system calls directly rather than using ptrace.
Dike says letting UML handle system calls directly will improve performance. It will be some time before UML can handle all system calls, but Dike predicts a "shrinking set of system calls that need to be ptraced, and a growing host of system calls that will run full speed."
Dike's talk Monday touched on the topic of management tools for UML and other virtualization software. Some of the attendees were keen to know whether more advanced management tools were planned for UML.
Dike says that he has "enough work to do in the kernel," and that he doesn't really think much about management tools or market concerns. However, he did say that the libvert project is "trying to make something under which everything that is known will fit under it. They're thinking about everything that's out there, including UML." Libvert is a C toolkit that started out as a way to work with Xen, but is also aiming to interact with other virtualization technologies under Linux.
I asked Dike about distributing software as virtual machine images. Dike says that this is and will continue to be a viable means for distributing things such as a system with a pre-installed LAMP stack or Snort, but it's unlikely to ever allow something like pre-made MythTV or game VMs, given the difficulty of getting the hardware to work.
Dike also noted that SWsoft and XenSource are trying to get OpenVZ and Xen technology, respectively, into the mainline kernel, but says that's unlikely. Dike says that Xen "doesn't fit in the Linux world," and called it "a technological dead end." He predicts a "family of virtualization stuff under KVM." Whatever makes it into the mainline kernel, says Dike, is what the distros will follow.
He also says that OpenVZ is unlikely to be adopted in the mainline kernel tree, at least as it is. Dike says that OpenVZ has to have "code sprinkled all over the place" to work, and it violates conventions within the kernel.
However, that's only the kernel world and technology adopted directly by Linux distros. In the market where people pick technologies based on different criteria, it's a different story. Outside the kernel world, Dike acknowledged that VMware is still far and away the leader in the virtualization market and says that the Linux KVM technology is unlikely to unseat VMware anytime soon. Not that he's worried about that sort of thing. "That's not what I think about, and I don't think most people in the kernel [development community] think about that.
"I keep my code clean, and get things knocked off my to-do list."