Virtually There: Steve Herrod's Blog

Wed, 16 Aug 2006

posted at: 14:26 | | | permanent link

Wed, 09 Aug 2006

Two of our engineers, Keith Adams and Ole Agesen, have a very interesting paper accepted at the upcoming ASPLOS XII conference. It's exciting to see that this conference starts with a dedicated virtualization session and that Keith and Ole's paper kicks off the whole thing.

Their paper, "A Comparison of Software and Hardware Techniques for x86 Virtualization," is the most detailed description of VMware's virtual machine monitor (vmm) available. It also examines the vmm implementation trade-offs as the CPU vendors directly support x86 virtualization with architectural extensions. You can download the paper here.

posted at: 19:43 | | | permanent link

Thu, 03 Aug 2006

There has been press coverage lately that plays up a growing battle between VMware and Xen. Following last week's story, "XenSource, VMware Conflict Holds Back Linux Virtualization", is another one published this week, "Oracle 'Losing Patience' with XenSource, VMware". Despite the headlines, the various groups focusing on Linux paravirtualization are actually working together well. Anthony Liguori (of the IBM Linux Technology Center) comments similarly in his follow to a recent Slashdot post:

I will say too that engineers from Xen and VMware have both been working together surprisingly well on this. There is an active conversation going on in the osdl-virtualization list and on other channels. Despite these stories, Xen and VMware are actually working together.

In the second article, Simon Crosby (CTO of XenSource) makes a similar observation:

The VMware team should be praised for engaging an open dialog with the Linux kernel and Xen communities, and they are actively engaging in the process.

I'd like to use this entry to give you a bit more detail about what's going and how we see this moving forward.

First and foremost, some have implied VMware is trying to slow down virtualization competition in the Linux space. This couldn't be further from the truth. In fact, we're extremely motivated to accelerate the industry to a generally accepted, technically sound, and open approach to paravirtualizing Linux as soon as possible. The motivation for proposing VMI as an alternate approach to Linux paravirtualization was to help ensure that Linux gets a sustainable, customer- and ISV-friendly set of interfaces. We think this is in the best long-term interest of the community. Obviously we're also quite interested in making sure that Linux's paravirtualization implementation is independent of any specific hypervisor implementation allowing VMware, Xen, and others to compete in this space. Virtualization competition is good for the customers and for Linux's continued growth.

Secondly, we've tried to be extremely open about our opinions and proposed approaches from the start. Our goal has always been a public debate focusing on the technical merits of different approaches. To recap, here's a timeline of VMware's activities in this area:

• July 2005 - We presented an initial version of the VMI "transparent virtualization" interface to key kernel and Xen developers at OLS 2005. This interface is now published on our website.
• August 2005 - We began to contribute Linux patches to the OSDL virtualization list with the goal of making Linux easier to virtualize.
• March 2006 - We released a public proposal and source code for the VMI paravirtualization solution on Linux. You can see these here and here. I also reiterated the specific goals of this approach in my May 4th blog.
• July 20, 2006 - We presented "VMI-Linux", a proposed implementation of the VMI spec on Linux, at OLS 2006, showing that this layer can be compatible with both VMware ESX Server and Xen. You can read the abstract or the paper itself on page 371 of the proceedings.

So we're trying to move as quickly as possible and in a very open way. where does that leave us today?

On the technical side, VMI-Linux is a working solution today for running Linux at high performance and on multiple hypervisors. It still needs some work though. For example, PAE and SMP support are available in the VMware ESX Server implementation, but are not yet completed on the Xen implementation. In other areas, VMI-Linux is fairly well polished. For example, it uses advanced inlining techniques to remove the execution overhead of indirect branches, preferentially inlining the native code and using direct branches to hypervisor-specific VMI functions. Since performance is the main goal of paravirtualization, this sort of optimization is quite useful. We are investing heavily in this area and continue to make improvements. And stay tuned ... we'll have more announcements shortly regarding some additional tools that help evaluate this general approach.

Overall, we believe the VMI-Linux is an excellent approach, but it's more important that the community as a whole quickly converges on an interface that maintains the main principles of the transparent paravirtualization approach. On this note, VMware, IBM, XenSource, and others are actively cooperating on a merged approach to kernel integration using an approach called "paravirt-ops". Rusty Russell is doing a great job of coordinating this effort. You can read more about paravirt-ops at his blog. And if you'd like to do real-time tracking of the collaboration, you can actually download and track the changes.

In the end, paravirt-ops may or may not end up being very similar to today's VMI-Linux approach. Regardless, we remain committed to working with the community to come up with a publicly designed Linux interface that works well for both existing and future hypervisors. It's great to see that the collaborating engineers are making daily progress on the paravirt-ops approach, which has the luxury of combining the best ideas from both the VMI and the XenLinux work!

So while there's certainly a ways to go, I do believe we're well on our way to an interface that is in the best interest of the Linux community and will let customers reap the great benefits of virtualization for years to come.

posted at: 17:27 | | | permanent link