Microsoft Speeds Up Its Data Center With Light and Mirrors


Microsoft Research is the first commercial customer of a new optical equipment module made by a seven-year-old startup that hopes its gear will enable servers to send and receive information faster. Lightfleet, based in Camas, Wash. sold an alpha version of its Direct Broadcast Optical Interconnect system, which uses broadcast light to connect computing nodes, to Microsoft’s eXtreme Computing Group, as part of a project to explore faster communication between servers in its cloud computing deployments.

Lightfleet’s gear looks pretty cool, and would help eliminate the bottleneck that occurs as information is sent inside servers and from server to server in dense computing environments. The company’s DBOI system uses light and mirrors to send bits from all compute nodes inside a server to all other nodes  at one time, rather than sending it via cables and a switch. So far, Lightfleet has raised $30 million in funding from angel investors and has 22 employees.

Other companies are addressing this bottleneck in a variety of ways that include using fiber in the data center, specialized gear or virtualizing the network fabric ( GigaOM Pro sub. req’d).  Intel (s intc) is proposing a similarly named optical cable technology called Light Peak for computer peripherals.

While Microsoft may have an industrial-scale strategy around its data center operations that seems antithetical to buying gear from startups,  its research arm shows that Redmond isn’t totally oblivious to new technologies to address the challenges of running hundreds of thousands of servers. Last week, I wrote about how Microsoft was looking for someone to work with solid state drives and ARM-based servers in its online services division.

Microsoft’s willingness to see the “light” when it comes to networking is just another example of how the shift to webscale computing may be opening opportunities for hardware and software startups, as the current generation of “commodity” x86 gear hits the wall. I will be leading a panel discussing  the prospects for hardware startups in a webscale world at our Structure 2010 conference in June, so we can see if this is the future or wishful thinking.


Stephen Johnson

server farms literally spend more money on electricity than they do on hardware. Also, the inefficiencies created at the switch (latency, contention, ETC) functionally cause the hardware in the various blades to function at a fraction of it’s theoretical capability. With the lightfleet architecture, this latency is dramatically reduced, not only because of the speed of optics, but also because it enables the first single-instruction one-to-all operation. What this translates into is that you can do the same computing with less hardware, equaling the same performance at roughly 1/3rd the power consumption and in a smaller physical footprint.


Stacey, what about Google? I read that they use very large quantities of cheap gear (which would mean commodity x86 machines) for their servers. Reliability is not an issue because when one machine fails the system shifts the work to another one. Surely Google needs massive processing of data as much as anyone, so are you sure x86 computers are hitting a wall for webscale computing?

Stacey Higginbotham

I am not. That’s why I think it’s such a rich topic of discussion. Clearly folks are pondering how to tweak the data center so it operates using less power while giving ever more performance and requires less human intervention. Does that open the door for alternatives? Several startups are trying to pitch that vision to me. Will the Microsoft’s and Googles of the world buy into it? I don’t know.

Comments are closed.