Google Warns Optical Networking Limitations Could Hinder Cloud Growth
Google exec details how the hyperscale network works today and why current optical technologies need to improve, dramatically.
At the core of all networks today are optical networks, which serve as the backbone of modern connectivity. Speaking in a keynote address at the OFC conference this week, Urs Hölzle, vice president of technical infrastructure at Google, detailed how Google has built its network and why it needs more improvement from optical vendors to help scale.
Hölzle explained that Google's cloud network started with a basic co-location approach. It then evolved to the current stage, which he referred to as Cloud 2.0 and is all about virtual machines and the services that run on top of them. What's just starting to evolve now is Cloud 3.0, which is another layer of abstraction and the beginning of serverless services.
"You don't even see the server or the network anymore," Hölzle said.
As the serverless network scales globally, Hölzle said that Google is now looking for step functions and 10x improvements in network capacity. He noted that if demand doubles every year, than 10x only provides a little more than three years of room for growth.
The current model for Google's networking is to use pluggable optics, which today are commercially available in 100 Gigabit Ethernet speeds.
"It's working but it (100 GbE) is also really bottlenecking what we do," Hölzle said. "Both the power and the cost of the solution is on the edge of what is possible."
"The density that you can get is already too low," he added.
Hölzle said that what Google wants is to move to some new form of optical networking module that is more compact, cheaper as well as something that is industrially manufactured.
"The optics industry today is still a bit of an artisan craft so to speak," Hölzle said. "If you really want to get 10x performance and get the cost to work, you have to automate the process."
As such, Hölzle commented that Google is looking forward to the day when optical networking components look much like other elements of data center computing that have become more of a commodity, following industrial automated manufacturing processing methods.
One particular area where Hölzle is eager to see innovation is in optical equipment for subsea cables. He noted that subsea cables today are very expensive, though they have robust capacity.
"What we really want are cheaper (subsea) cables, with perhaps lower capacity, but we'd have multiple cables that are physically separated," Hölzle said. "With every additional cable we'd exponentially decrease the failure amount."
"So if we could buy it, we'd rather have 30 cables across the Pacific and not three," Hölzle continued. "It would be a better solution for us."
Hölzle also wants to see continued programmability and flattening across both the IP and optical layer of the network to enable improved manageability and agility for application deployment. Overall Hölzle said that cloud architecture increases bandwidth demands, which is why more innovation is needed at the optical layer to dramatically help hyperscale networks to grow.
"Thing have been going well in the last few years, but we're rapidly approaching a wall," Hölzle said. "We have to find a way take off and skip that wall rather than crashing into it."
Sean Michael Kerner is a senior editor at EnterpriseNetworkingPlanet and InternetNews.com. Follow him on Twitter @TechJournalist.