AMD Shanghai chips target virtualization performance

AMD's new Shanghai processors are designed to improve performance of applications running in virtual machines with various features that reduce the amount of work required from a hypervisor.

"The idea is to decrease the overhead associated with a virtualized application, and make it run as near native as possible," says Margaret Lewis, director of AMD's commercial solutions.

Both AMD and Intel began embedding virtualization-friendly features in their chips a few years ago, allowing routine operations to be handled in the hardware and reducing the amount of software code needed from hypervisors such as VMware and Xen.

AMD Thursday launched its latest quad-core Opteron processor, code-named Shanghai. Lewis says virtualization improvements will result in benefits such as applications being able to handle more users or perform more transactions. AMD has fine-tuned its Rapid Virtualization Indexing technology, which reduces virtualization-related performance hits by taking functionality typically performed within software and moving it to the CPU. The goal is "near-real time application performance" that's particularly important for transaction-intensive applications such as Web serving and databases, AMD says.

AMD did some tuning of its world switch, the technology that enables the switching from one guest operating system instance to another, Lewis says. AMD also targeted the translation lookaside buffer (TLB), a CPU cache, with an innovation that lets the hypervisor move from one task to another more quickly.

Typically, a virtualized application's memory has to be flushed each time the hypervisor moves from one virtual machine to another in order to perform various tasks, Lewis explains. AMD enabled the tagging of memory so the entire memory buffer doesn't have to be flushed. Memory can thus be reloaded fast and efficiently when the hypervisor returns to the original virtual machine.

"There are parts of the buffer that don't need to be flushed," Lewis says. "By tagging them, the virtualization software can decide what needs to be flushed and what doesn't."

AMD plans more virtualization-related improvements in future processors. Lewis notes that enterprises are becoming more interested in streaming desktops from servers to clients. "We're starting discussions about how to handle graphics in those environments," she says. (Compare desktop and client management products.)

AMD will follow Shanghai with Istanbul, a six-core processor in 2009, followed by the 12-core Magny-Cours in 2010, the IDG News Service reports.