With lasers, almost time to cut the data center cord

Wire report: Thick tangles of computer cables (and the high costs of maintaining them) may soon be history at the nation's many data centers.
By GREGORY ZELLER //

From the Credit Where Credit Is Due file – and the Stony Brook University Department of Computer Science – comes a potentially groundbreaking innovation laser-focused on the future of data centers.

The plan: to use light beams and mirrors to replace the thousands of miles of cables snaking through the nation’s digital hubs, creating quantum leaps in performance and significant energy-usage reductions.

But while recent coverage of a National Science Foundation-funded optics-infrastructure project by Stony Brook, Carnegie Mellon and Penn State universities has centered on Penn State’s contributions – even the February newsletter from SBU’s own Center of Excellence in Wireless and Information Technology reprinted a Technology.org article trumpeting the Nittany Lion angle – the idea and the project were both born on Long Island.

Himanshu Gupta: Data driven.

So says Himanshu Gupta, an SBU associate professor of computer science who is heading up the three-headed project – which stems from an idea originally conceived by “one of the brightest students I’ve ever had,” Gupta told Innovate LI.

“We were thinking about wireless architecture for data centers and he suggested the idea of laser communications,” Gupta said, noting he, SBU associate mechanical engineering professor Jon Longtin, SBU College of Engineering and Applied Sciences Associate Dean Samir Das and Vyas Sekar – a former SBU researcher now teaching electrical and computer engineering at Carnegie Mellon – first “put our heads together” on the student’s big idea in 2013.

“We wrote, basically, a small paper that had the basic idea,” Gupta added. “We expanded the idea to include the entire architecture and the ability to handle all the issues and challenges in a larger paper in 2014.”

It was subsequent to that, according to Gupta, that Mohsen Kavehrad, a highly credentialed Penn State electrical engineering professor and featured performer in the Technology.org article, became involved.

“The initial motivation to involve him was to write a National Science Foundation proposal together,” Gupta said. “Then it got funded and we started doing some work together.”

The SBU scientist is by no means downplaying the contributions of Kavehrad and his team. Researchers at all three schools “talk and collaborate and share ideas,” Gupta noted, precisely the sort of all-hands teamwork required of a project with such far-reaching potential.

Data centers – secured, environment-controlled storage spaces for computer network servers – are the nerve points of most modern information systems. Because of their piecemeal nature, they’re often constructed and expanded a bit at a time, leading to a ragtag assortment of computer racks and other equipment. And the servers and other computers inside the large spaces tend to run warm, meaning large data-center spaces must be air conditioned, usually well below normal human comfort levels.

Of course, even as computer science wades deeper into the fiber-optics age, the centers also pack incalculable lengths of hard wires – making images of precious data center computer cables twisted and knotted like overcooked spaghetti fairly common.

While comical in appearance, the techno-tangle is no laughing matter to engineers charged with maintaining data center efficiency – or to anyone who has to pay the electric bills that keep centers running, and keep them cooled.

Mirror, mirror: Beaming up the data. (Photo: Penn State University)

Enter the Free-Space Optical Inter-Rack Network with High Flexibility (stretching the entomology and purpose of acronyms, its creators came up with “Firefly”), the heart of the NSF-funded project. In a nutshell, Firefly uses infrared lasers and receivers to transmit information, eliminating the need for solid cables.

There are several inherent challenges, Gupta noted, including the ability to “be able to reconfigure these laser links on the fly.” The laser receivers must also be very small – most data-center racks are just 2 feet by 4 feet, according to Gupta, and “you need about 50 of these link terminals on top of each rack.”

Another hurdle involves configuration – that is, arranging the multiple laser beams so that links can be steered in real time “with high precision,” Gupta said, noting an original design incorporating a ceiling-mounted, beam-refracting mirror proved impractical and had to be scrapped.

Now, project leaders are focusing on rack vibrations caused by the servers’ internal fans and motors. The shaking is minuscule – on the order of millimeters, according to Gupta – but can have a major effect on laser-based communications.

“Because of these vibrations, the links can get lost,” he noted. “This technology requires extreme precision, and we are looking at how we can make up for these vibrations using real-time tracking and pointing.”

The project’s ultimate goal is to “automate the process,” Gupta said, meaning instead of 50 or so link terminals, outfitting each server rack with a single board that provides real-time alignment, minimizes maintenance and bottoms out data-center energy costs, with less electricity needed to run the systems and keep the environment temperature controlled.

“That’s the holy grail,” he noted. “That’s a technology with commercialization potential.”

That’s also a ways off, Gupta added, and for now, the important thing is that all project participants continue their experiments with microelectromechanical systems, tiny targeting mirrors and other cutting-edge tech.

Based on current science and well into the foreseeable future, hard wires will always be the most efficient way to transmit data, according to Gupta, who admits that “if you have enough money to put in enough wires, you cannot do better than that.”

But consider data centers’ growing monetary and energy costs and the wireless benefits come clear. In the Technology.org article, Kavehrad estimates that by 2020, national data hubs will use a total of 140 billion kilowatts of electricity per hour – about $13 billion worth by today’s rates, or the equivalent output of 50 average power plants.

Wireless data centers can happen, Gupta said, but only if the partners at the three universities continue to work together – and some new benefactor “shows external interest” once the NSF funding runs out.

“We have a vision, and it’s only a few years away,” Gupta added. “The technology is not very difficult. It’s just a matter of interest and effort.”


Comments are closed.