CHICAGO -- Light Reading's Big Telecom Event -- Google is opening the network configuration data and network topology models that it's been developing for years, publishing the model for comment and contributions by others.
Speaking in a keynote presentation here and in an interview afterward, Bikash Koley, principal architect and manager of network architecture at Google (Nasdaq: GOOG), said the Internet firm realized it needed the input of other large network operators to define vendor-neutral network data and network configuration models that will enable a "true software-defined network model" and a better Internet.
"We have been working on this for many years, but we realized that, if we have any hope of having a common description for networks and equipment, we need to open this up," Koley said in an interview after his speech. "We want to jump start this process, to get it moving, and we are absolutely open to working with other large carriers to get this done."
To that end, Google will make its models public, most likely this summer, and open for discussion in standards bodies, most likely starting with the Internet Engineering Task Force (IETF) . The models are based on protocol buffer, a structured data format developed by Google that has an open format and has been published for general use.
The industry, Koley explained, talks a lot about the network data plane and the network control plane, but it tends to hand wave over the management plane in SDN. The management plane is extremely important, however, because it defines how services and applications are orchestrated. "In order to run a large infrastructure, you need abstraction. In order to implement abstraction, you need models," said Koley. "Those models don't exist in the network layer today." The network configuration and network topology models are essential to Google in its efforts to build multi-layer, vendor-neutral networks that allow carriers who swap traffic to do it more optimally and improve the end-user experience.
"We all benefit from a better Internet," he said.
In his keynote, Koley spoke in a rapid-fire keynote about SDN, or "as we call it at Google, just networking." The company's experience comes from rolling out its own wide-area software-defined network known as B4, a network that went into production in early 2011. (See How Do You Get to SDN From Here?.)
Through the B4 deployment process, Google discovered both the significant advantages of SDN – like the benefits of having a common network operating system and common network APIs – as well as the many areas where major technology gaps still exist.
Google has made limited progress in the management plane, where Netconf is a promising option but needs universal adoption, and very limited progress toward the standard network data model. In this arena, Yang as a data modeling language holds promise, but there still needs to be vendor-neutral ways to describe network/device configuration, he said, along with a standard network topology model.
In today's network, the network state is detected by information provided by routing devices, but those are based on vendor proprietary approaches. What's lacking is a vendor-neutral network model that also covers all the layers of the network, including legacy devices.
"The second thing we discovered is that, if you try to build a model which is multi-layer and scales to a global network with thousands of end points, you need a data structure and messaging format that is extremely optimized," he said. "We have tried doing it in Yang and UML [unified modeling language], but those are bulky, and it's hard to describe."
Ultimately, any device built for the network must match the network model and be standard, Koley said, in the same way that, in today's Windows world, devices built to predefined network and data models and can be plugged into a system easily and work from day one as expected.
One major benefit to major network operators of establishing these network configuration data and topology models is that it will make it easier for them to abstract their legacy network systems as part of the process of moving to virtualized networks. "We want a model that can describe anything from layer 0 to layer 7," he said.