Facebook has some ideas for automating the management of the optical network, but it's a trickier task than automating packet networks, because of the differences between vendors' implmentations.
Still, it's doable, and the key is the mixture of an abstraction layer and a lot of Facebook programming knowhow, said Sri Bala, a Facebook principal architect. He presented the company's ideas, most of which are still in the lab, last week during the Next Generation Optical Networking conference, held in Burlingame, Calif.
Some of this work could be taken open source, but so far, Facebook is doing it all solo. "We have a lot of internal tools, which gives us sufficient compute power," Bala said -- quite the understatement, given that Facebook has enough compute power to run, you know, Facebook.
The idea behind Facebook's automation work was to get engineers away from maintaining the optical network. "At Facebook, we believe in engineers building robots. Robots manage then network," Bala said.
The key is abstraction -- masking the differences between vendors' gear by using a common abstraction layer to talk to all of them. Whoever or whatever is making decisions, be it a human operator or, say, an SDN controller, would talk only to the abstraction layer, while software in that layer would send translated commands to the network management systems (NMS) in various pieces of gear.
Unlike packet networks, optical networks have to take into consideration things like wavelength assignments and optical power levels. Vendors handle these issues differently, so it's been easiest to use one vendor's gear for both sides of an optical link. Any changes to the network -- even at Facebook -- are still highly manual, Bala said.
So, Facebook is developing an NMS Adaptive Layer, which would provide that abstraction layer to talk to these different vendors' NMSs, taking advantage of the increasing use of APIs in network management.
Bala also sketched out ideas for automating some elements of decision-making. Facebook has written a Spectrum Assignment Engine, for example, which would make wavelength assignments while taking forecasted demand into account. That information would be fed through an orchestrator down to the NMS Adaptive Layer.
"The time from demands to bits is going to be much faster" that way, he said.
To truly make this work, Facebook would also have to gather telemetry from optical networking gear, using the information to make decisions about how to optimize the network. For example, deciding if, during a failover situation, an optical connection could swap to a route that has a different modulation scheme, just temporarily.
— Craig Matsumoto, Editor-in-Chief, Light Reading