Facebook Redefines Data Centers, Parallel with Intel Photonics Share your comment!

The world’s largest datacenters at Facebook will leapfrog competition by redistributing parallel system resources into homogeneous racks, and use Intel’s next-generation silicon photonics to interconnect Xeon and Atom processor racks to into more energy-efficient, and less expensive storage and networking racks.

“Intel is collaborating with Facebook to redefine the future of mega-datacenters,” says Justin Rattner, Intel’s chief technology officer (CTO). “By using Intel’s new 100-Gbit per second (bps) silicon photonics technology, compute, network and storage systems can increase bandwidth, with fewer cables and greater power efficiency, compared to today’s copper based interconnects.”

Today, each server has its compute, networking and storage resources aggregated in the same rack to reduce the distance copper interconnections must travel between them. In contrast, next generation mega-datacenters will disaggregate resources into separate compute, networking and storage racks. These are interconnected by Intel’s 100-Gbit per second optics — at a fraction of today’s cost and power consumption.

 At the Open Computer Summit 2013 (Jan. 16-17, Santa Clara, Calif.) Quanta Computer unveiled an early prototype of Facebook’s new disaggregated rack architecture. The design used distributed 100Gbps input/output Ethernet switches connecting arrays of Xeon and ultra-low-power 22-nanometer AtomAvoton” system-on-chip (SoC) processors to networking racks and server-farms. 

Supercomputer Interconnection Fabrics Based on Silicon

For the past decade, Intel has been developing all the necessary silicon-photonic components to bring down the cost of supercomputer-caliber fiber-optical interconnection fabrics. Doing permits the selection of architectural features based on performance and energy efficiency, rather than on minimizing the amount of copper used in the interconnect. Intel now has emitters (lasers), electrical-to-optical converters (modulators) and detectors (photo diodes) all made from inexpensive silicon, rather that the exotics gallium-arsenide compounds used in supercomputer-caliber interconnection fabrics today.

After developing the separate components, Intel has spent the last two years perfecting the architecture so it can handle the entire interconnection task with inexpensive silicon components. The company is now delivering engineering samples of its new photonics receptacle to the Open Computer Foundation, in cooperation with optical-fiber maker Corning and the owner of the world’s first photonic-enabled mega-datacenter, Facebook. 

Open Computer Project (OCP) Seeks Standardization

“Developing in the open and contributing to the Open Compute Project will accelerate the pace of innovation,” says Frank Frankovsky, vice president of hardware design at Facebook and chairman of the Open Compute Foundation. “With these technologies, the entire industry can close the utilization gap that exists in today’s system designs.”

Besides lowering the cost of ownership, extending the lifetime of each type of resource–since they can be upgraded separately–disaggregated mega-datacenters also eliminate performance bottlenecks while shrinking the datacenter’s footprint and lowering its power consumption, according to Frankovsky.

 Intel and Facebook are both founding members of the Open Computer Project (OCP) and are working together with others to standardize OCP boards for Xeon and Atom processors as compute resources, smart-storage resources and other devices capable of using Intel’s photonic receptacle.

Posted on January 22, 2013 by R. Colin Johnson, Geeknet Contributing Editor