Barcelona: Quad-Core Opterons Now Feature Virtualization Support

The latest batch of 8300 series and 2300 series Quad-Core Opteron processors, formally announced by AMD today, will feature hardware-based library support for software to be able to run virtualized environments.

The AMD-V virtualization layer, as the company will call it, will provide virtual machine environments from VMware, Microsoft, Xen, and others with direct channels to memory and system resources, enabling them to bypass operating system drivers and go straight to the source. The result is something almost, but not entirely, quite unlike virtualization as we have come to know it before. Typically, a virtual machine establishes an intentional layer of indirection between a guest environment and system hardware. Now, these systems will have the option of addressing memory directly, rather than indirectly.

BetaNews spoke at length with John Fruehe, AMD's worldwide market development manager for server and workstation products. During that discussion, Fruehe (pronounced "free") explained the key distinctions between his company's concept of virtualization support and his competitor's.

SCOTT FULTON, BetaNews: Intel has vPro technology...and its idea of virtualization for the business platform is to be able to support a hardware-driven hypervisor that runs an operating system on a completely virtualized scale. I take it AMD-V is something different than that.

JOHN FRUEHE, Worldwide Mktg. Dev. Mgr., AMD: Yea, we've always had the advantage in virtualization, because if you look at its profile, the virtualization platform is so memory dependent, and memory and I/O are really the key pieces for virtualization. We've always had the advantage there, so as far as we're concerned, we've always been in really good shape from a virtualization perspective.

What AMD-V does is, it really changes virtualization and makes it even more memory efficient. I think my competitor's view is, "Let's make that hypervisor layer as fat as possible, and put all the functionality there." And as far as AMD is concerned, we believe we're much better off to put a lot of that functionality into the silicon itself, and let it run natively in hardware. One of the things we do with Barcelona is, we add a functionality called Rapid Virtualization Indexing. To give you an idea of how this works, Rapid Virtualization Indexing allows the virtual machine to talk directly to the memory instead of having to do all of those memory translations through the hypervisor layers. So all of that memory translation happens in hardware and silicon instead of happening in software. So you get much better performance and scalability.

In some previous benchmarks we were able to do with VMware, we actually dropped the memory overhead by about 70 or 80%. To give you an idea of how this works, last two weeks, I was in Asia and talking a lot about virtualization with the press there, and I'm not smart enough to be able to speak Mandarin, so when I'm talking to the press in Beijing or in Taiwan, I've got to have a translator. So I say something in English, they translate it into Mandarin and the reporter understands it, the reporter asks a question in Mandarin, it goes back to the translator, she turns it over into English back to me and I can understand it.

Now, if I was smart enough to be able to speak Mandarin, I could have a one-on-one conversation with a reporter, and it would go a lot like the conversation we're having. What Rapid Virtualization Indexing does is really remove a lot of that translation from having to happen at the slow software layer, allowing it to happen at the hardware layer so it happens significantly faster, much like if I was able to speak Mandarin and you see significant performance increases.

The other advantage, from a virtualization perspective, is [that] it is so memory dependent, and because I have an integrated memory controller, the number of virtual machines sharing a memory controller is significantly reduced. So you tend to get a much better ratio of virtual machines to memory controllers, and most customers like to do virtualization on four-way platforms.

So in a four-way platform, I've got four memory controllers, one per processor, and so if you were running 16 virtual machines, for instance, on a four-way server, in my environment, I've got four virtual machines per memory controller, and in the Intel world, you've got 16 virtual machines all vying for attention from that single memory controller. And you can start to see how, from a scalability standpoint, you can get much better performance out of an AMD platform because there is so much lower overhead for each memory controller, and the AMD-V makes it that much better.

Next: In the immortal words of Lawrence Olivier, "Is it safe?"