To keep up with the growing demand from big data centres, optical technologies need to take significant steps forward, according to a Google executive.

Urs Hölzle, vice president of technical infrastructure at Google, called for 2x cost reductions, 10x capacity gains, and more programmability in a keynote at the annual OFC event but gave few specifics on how to get there.

The talk once again underscored how big data centres are pushing the limits of communications technology. Ironically, the web giants are extreme both in their growth and in their sensitivity to cost, putting pressure on chip, component and system makers.

“Things have been great for the last five years, but we’re rapidly approaching a wall, so we’re looking for step-function changes, not incremental changes,” Hölzle said.

“We see 2x growth in nearly all the [optical] markets—that’s true in the data centre, in campus/metro networks where we connect data centres for storage replication, and to get to the Net and in the WAN where we deliver apps,” he said.

The demand is partly driven by what Hölzle called a new generation of cloud services that Google is offering based on the latest level of virtualisation.

“You never see a specific server or disk; you just provide data and queries. We pick the number and kind of machines, so you don’t have to worry about machine details,” he said. “To keep the illusion of server-less computing, we need high-performance, high-capacity networks.”

Specifically, Hölzle called for more simplicity through automation in the design of optical modules, lower-cost undersea cables with greater redundancy and support for programmable control planes that span optical and copper nets.

“The optics industry today is still more of an artisan craft,” he said, noting pluggable modules that look like USB sticks.

Each network switch may use dozens of these modules and Google probably owns thousands of such switches. While not providing any specifics, he called for module makers to take a lesson from the integration and automation of the semiconductor industry.

“The existing way of terminating fiber doesn’t work. Please build me a robot that terminates fibre—and it only has to be right 99% of the time,” he said.

More programmable networks win Today’s submarine cables pack as much as 144Tbits/s in coherent optics, but they are expensive and unreliable. “They don’t even give you three-nines reliability and can’t even protect against the anchor of the ship” that lays them in much the same way as the first undersea cables in the 1860s, he said.

Hölzle called for cheaper bundles that might use 30 rather than three cables to provide more redundancy. Google is now laying its third submarine cable across the Pacific Ocean.

In another sign of how big data centres are driving comms, one OFC paper described a trial by Facebook and Nokia Bell Labs using a trans-Atlantic cable. It used new probabilistic constellation-shaping technology and shaped 64-QAM over a 5,500-km cable between New York and Ireland to increase capacity almost 2.5 times, delivering spectral efficiency of 7.46b/s/Hz.

Finally, Google is trying to control all of its networks using mainly traditional programming tools on standard servers. The approach helps it balance workloads and handle failures across its geographically dispersed data centres. “The market for hyperscalers will go to the network that is most programmable,” he said.

However, Google creates its own network control plane that oversees its copper and optical nets. Thus, vendors need to figure out how to work with it rather than create something new.

“We can’t use your beautiful control plane that is optical only,” he said.

In the end, “I hope you want to lower costs 2x every year and increase capacity 10x and make it all programmable,” he told OFC attendees.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved