Facebook Moves to Crush Servers in a Group Hug
Facebook released a Common Slot architecture specification for data center motherboards
By: Maureen O'Gara
Jan. 21, 2013 08:00 AM
Um, something happened this week you ought to know about.
Facebook blew up the traditional monolithic server - and lit charges under the entire $55 billion-a-year server industry.
GigaOm was first to say it that way and it may turn out to be true so it bears repeating.
Facebook, along with its user-leaning Open Compute contingent, is bent on redesigning servers to suit themselves using interchangeable, disaggregated, independently upgradeable parts.
Ultimately it's supposed to free the customer from the tyranny of the vendor roadmap.
To advance this crusade, Facebook released a Common Slot architecture specification for data center motherboards at the Open Compute Summit Wednesday.
The thing is nicknamed "Group Hug" and it's supposed to produce boards that are completely vendor-neutral and last through multiple generations of processors from multiple vendors.
Having been born too late to exert any influence over server blades, Facebook is determined to see that the new microserver architectures conform to some sort of compatibility code.
Intel, AMD, Applied Micro and Calxeda are already committed to producing products designed to the Common Slot spec and Calxeda, the little Texas start-up with the ARM microserver designs, is so pleased to be in this rarified company it's beside itself.
The way things are unfolding it looks like Facebook and the Open Compute Project (OPC) are endorsing microservers built out of the mobile chips used in smartphones and tablets, giving ARM a chance to break into the citadel held fast by the x86.
Frank Frankovsky, Facebook's VP of hardware design and supply chain as well as executive director of the Open Compute Foundation, showed off a Group Hug board with five unreleased Intel x86 Avoton S Series Atom chips on it as well as five X-Gene 64-bit ARM SoCs.
They all share the same power, electrical and mechanical interconnects and slide into the same microserver chassis.
The chips are on cards that are inserted into the so-called common slot. The motherboard can currently accommodate 10 cards.
"We're establishing for the first time a common slot for any SoC maker to design to a common standard," Frankovsky said. "All the surrounding bits are the same, with DDR memory and network controllers, and now for the first time we will have the ability to have a common slot architecture."
It uses a simple PCIe x8 connector to link the SoCs to the board.
"If we had left this to the industry they probably would have gone out and found the most expensive and esoteric connector on the planet," he said. "What we decided to do was use a PCI-e x8 connector and simply change the pin-out." The Facebook backplane design has one PCI pin-out per server.
The board's layout isn't etched in stone either. It's just supposed to encourage people to use the common slot for their CPUs. "We will now not be bound by placement of components on a single monolithic motherboard," Frankovsky said. "We will be able to do smarter tech refreshes."
The object of the game is to make hardware that's cheaper, greener, more upgradeable and software-defined so it fits the workload and less power hungry.
Group Hug envisions abandoning the modern vendor-integrated server that has to be switched out generation-to-generation for components that can be upgraded as they become available without scraping what surrounds them, letting customers design modular, custom, scalable servers with just the right compute, storage and networking for the job.
The growing consensus is you shouldn't have to change the whole system just to refresh processors, memory or I/O.
The concept and the movement building behind it obviously threaten IBM, HP, Dell and probably VMware too, since Facebook doesn't much fancy virtualization as a way to drive hardware utilization.
Intel, on the other hand, is on the movement's board and is contributing designs for its forthcoming silicon photonics technology, which will enable 100 Gbps interconnects, enough bandwidth to serve multiple processor generations.
Frankovsky said, "This technology also has such low latency that we can take components that previously needed to be bound to the same motherboard and begin to spread them out within a rack."
"We'll said be able to do things in the data center that we've never been able to do before," Intel CTO Justin Rattner said.
It's supposed to connect servers together using a laser-base technology created using silicon rather than pricier techniques.
Reader Feedback: Page 1 of 1
Latest Cloud Developer Stories
Subscribe to the World's Most Powerful Newsletters
Subscribe to Our Rss Feeds & Get Your SYS-CON News Live!
SYS-CON Featured Whitepapers
Most Read This Week