The Digital Edge
|Back to Page One|
|Rick Merritt||April 19th 2012|
A Google executive gave a rare peek inside the Web giant’s data center networks to show the OpenFlow standard it backs for software-defined networks is ready for commercial use. Google is using OpenFlow on custom-designed hardware for all the internal networks it runs connecting its global data centers, said Urs Holzle, senior vice president of technology infrastructure at Google, speaking in a keynote at the second annual Open Networking Summit here.
OpenFlow is a technique for controlling network operations in software run on centralized computer servers saving cost, time and power. It aims to simplify and virtualize today’s business networks that currently require a number of specialized, distributed systems, each with its own software load.
If OpenFlow becomes widely adopted, it could disrupt the fortunes of major router and switch makers such as Alcatel-Lucent and Cisco Systems as well as the ASICs and embedded processors they use. Google has enable “centralized traffic engineering” on its network using OpenFlow. So far it has found it can run such functions “literally 25 to 50 times faster on a 32-core workstation,” Holzle said.
“It becomes easy to do things that are hard to do on embedded processors typically with little memory on a networking box,” he said. “You can use all the [computer] tools for normal software development, and that makes it faster to develop software that is higher in quality,” he added.
In 2009, Google started testing OpenFlow code from Stanford’s Clean Slate project before the software became an official standard. It now uses OpenFlow as the basis for its so-called G-Scale network that links its global data centers. G-Scale actually carries more traffic than a separate Google network that serves its external end users. “I didn’t expect 18 months after we started tests we could really carry all our [G-Scale] production traffic” on OpenFlow, said Holzle.
The network is running on custom 10 Gbit/second switches with 128 ports Google built from standard merchant chip sets. Holzle did not detail the internals of the design. Functionally, the Google OpenFlow switch “runs almost no software, just the OpenFlow agent” using just BGP and ISIS protocols, Holzle said. “We wanted to see how far we could go moving software off the box. The hardware is a side piece we had to do. I would love to be able to buy this, and I am confident I can get such systems this year or next,” he said.
In a separate conversation after the keynote, Holzle said Google does not expect to buy OpenFlow systems this year as it focuses on finishing the implementation of its current G-Scale network. However he opened the door to purchases in 2013 and beyond, probably looking for 40G systems supporting as many as 1,000 ports.
Google OpenFlow WAN Diagram; Usage Chart
Google is still rolling out new features and procedures it wants to use on the OpenFlow network. Once they are running it will begin to quantify specific cost savings it already anticipates. “We are clearly ahead in total cost of ownership and amortizing the development of [the switches],” Holzle said. Network “utilization improvements are clear, we are already at the same or better guarantees of service and once the network is fully implemented we expect we can operate and the same or higher levels with less effort,” he said.
In particular, OpenFlow promises to significantly reduce the time administrators currently spend configuring specialized systems, rolling out new network features and responding to outages. “Instead of updating hundreds of systems with thin memory and a little CPU, you update a few powerful control servers with tons of memory,” said Holzle. “I am highly confident we will see substantial reductions in unit cost per month at given service levels, and we are already seeing signs of improvements better than any other technology change we might expect,” he said.
At this stage the OpenFlow network is robust and secure, but challenges are still ahead. It is particularly difficult to determine how to coordinate redundant OpenFlow controllers, Holzle said. In addition, the standard is still in an early stage. The current spec doesn’t take advantage of features embedded in the ASICs in today’s routers and switches, according to engineers working on the next version of the standard.
In terms of the standard’s robustness, “I was willing to have more outages than we did,” Holzle said. In addition, “there’s no log-in to the box anymore, and intrinsically that makes you net more secure,” he said.
Google is acknowledged as running what is currently the world’s largest production OpenFlow network. But the technology is also at use in other business networks. In a separate talk, an NEC executive said Genesis Hosting Solutions saw a 60 percent reduction in the number of global IP addresses it needed to support by moving to OpenFlow. Nippon Express Corp. has reduced rack space requirements, power needs and downtime by 70 percent or more with an OpenFlow network, he added.
Nearly 900 people attended the OpenFlow event including many executives from service providers, large switch and router makers and chip companies. The OpenFlow Foundation now has more than 65 members.
Rick Merritt writes for EE Times, from where this article is adapted.