Dell skunkworks brews ARM server future

Cortex A9. LAMP stack. Eye of newt

Chip Marvell

Chip maker Marvell, which bought Intel's XScale ARM RISC business five years ago, is hoping to get its dual-core variant of the Cortex-A9 chip into the field by the end of the year. As El Reg has previously reported, a startup called Smooth-Stone has raised $48m in funding to design ARM-based servers.

In the summer, there were rumors Facebook was putting ARM-based servers into its data center in Oregon, and of course it would be interesting to see Apple deploy ARM-based servers in its massive North Carolina data center.

Prince and his colleagues at Dell are going to be putting Cortex-A9 skunkworks servers together as soon as chips and motherboards are ready, and will load them up with a LAMP stack to get the feeds and speeds of what these systems can do.

So when the next hyperscale data center looking for bespoke servers to radically cut back on power consumption comes knocking on Dell's door, the company will have another option beside Intel Atom and low-voltage Xeon, Opteron and low-voltage Athlon, and VIA Technologies Nano chips.

Dell is not expecting a lot of business for such machines, but it has to be ready to capture whatever there will be. That is the secret of the success of the Data Center Solutions unit of Dell, which has quietly become a dominant force in Dell's overall server shipments.

"ARM is moving the core forward with the Cortex-A15, and the ecosystem is, too," says Prince, adding that it the ecosystem is what will add the memory capacity, I/O features, and other functions to an ARM chip design that make it suitable for a server. "This feels like it is moving at the right pace."

Virtualization of memory

Some would argue that the ARM chips are not evolving fast enough to take on Intel and AMD in the server racket. As El Reg explained in August, the Cortex-A15 design includes a feature called large physical address extension, which will translate 32-bit virtual memory addresses to 40-bit or larger physical addresses.

This gives hypervisors and operating systems running on the Eagle machines some sleight of hand to address more physical memory than the 32-bit limit allows (at 4 GB). (Similar virtualization of memory was used in 16-bit and then 32-bit PC server chips as Intel moved from the desktop to the servers, so there is some precedence for this.)

ARM's own specs for the Cortex-A15 calls for "home and Web 2.0 servers" in quad-core configurations with the Eagle chips running at between 1.5 GHz and 2.5 GHz and with virtual machines atop the cores and support for more than 4 GB of main memory in the box.

Don't get the wrong idea. We are not going to see Dell launch a PowerEdge-A series lineup until ARM chips go mainstream in servers.

"I don't see ARM taking over the enterprise server market," says Prince. "But ARM could be a very important niche."

Ubuntu Linux and Android Linux already run on ARM chips, and so does Microsoft's Windows Embedded CE. There is no reason why the real server versions of any of these platforms - and others - can't be tweaked to run on ARM-based servers.

Lots of hyperscale customers control their own code-bases, and porting to a new architecture is no big deal for their techies. If the cost per unit of performance per watt numbers work out, it is cheaper to pay a few programmers to retune the apps for a new architecture than it is to build a new data center or three.

Dell did not provide access to its benchmark results on the LAMP stack running on the skunkworks Cortex-A8 machines, but you can see a performance test done by ARM Holdings that pits a netbook using a 1.6 GHz Atom processor against a prototype machine using a dual-core Cortex-A9 processors here.

See if you can tell if there is much of a difference between the two. ®