I spent a couple of days last week at the Open Networking Summit in Santa Clara. The conference covered a wide range of business and technical subjects relating to Software Defined Networks, with the topic of OpenFlow receiving particular attention.
In a keynote address on the first day of the conference, Urs Hoelzle, Senior Vice President at Google, described how Google used OpenFlow as part of a major overhaul of its internal networking infrastructure to reduce costs and increase efficiency. This internal “G-scale” network connecting Google’s vast data centers worldwide actually carries more traffic than its external “I-scale” customer-facing network.
Hoelzle presented a fascinating description of how Google approached the massive task of swapping their network hardware while, of course, ensuring no disruption to on-going traffic. Having concluded when they started the project that there was no off-the-shelf networking equipment “even remotely suitable for this task”, Google developed their own 128x10G chassis using merchant silicon, ready for deployment starting in early 2010. They then installed this in each data center, using a method whereby they first pre-deployed the equipment at a site, taking down half the site’s networking equipment and connecting it to the new system. After testing to verify that the upgrade worked, the engineers would then repeat the process for the remaining 50 percent of the site’s equipment. By early this year, Google’s entire internal network was running the new, OpenFlow-based equipment.
Google now has a central traffic engineering controller that optimizes routing and provides more predictable performance. OpenFlow allows network packets to be routed by software running on routers. By separating packet switching from network management, Google has better and easier control over their network.
Many speakers pointed to Google’s announcement as a key proof point for OpenFlow, demonstrating that the technology is usable in the world’s largest networks and delivers quantifiable business benefits. This view was certainly reinforced by other case studies presented at the conference, by organizations as diverse as Kindred Healthcare, Indiana University and NTT Communications.
One of the questions debated at the conference was how quickly OpenFlow will gain traction in broader enterprise and data center applications. A couple of speakers pointed out that Google of course has access to resources way beyond what’s available to most IT organizations, while service providers are naturally hesitant to introduce new technology into something as complex as the average enterprise network.
Martin Casado, Nicira’s CTO, talked about the value of OpenFlow within virtual switches (announcing that the OpenFlow Virtual Switch is now included in the Linux kernel), but recommended using Layer 3 to manage the external fabric interconnect. Speakers from Dell, HP and other companies mentioned that merchant silicon (e.g. switches) is not yet available with fully optimizations for OpenFlow.
A common thread within the presentations was the likely coexistence between existing networks and SDN. Speakers from Cisco and Juniper both talked about this, explaining that they expect to leverage traditional routing protocols with SDNs for the foreseeable future. Cisco described how SDNs should be linked to analytics and policies, using visibility into the control, data, management and transport planes to drive more intelligent decisions.
Finally, there were some interesting discussions about what “killer apps” might emerge as the driving force behind Software Defined Networking and OpenFlow. Network virtualization was mentioned, as was fine-grained policy for Bring Your Own Device scenarios. A couple of people talked about a future with “app stores” providing off-the-shelf networking applications that would (presumably) run on standard, secure, risk-free networking APIs. Personally, I would settle for not having to reboot my home router every time there’s a full moon, or a decent GUI for updating its firmware.
What are your thoughts on this topic? What do you expect to happen in terms of OpenFlow’s adoption in enterprise networks and data centers? What are the key issues (both business and technical) that the industry needs to resolve to accelerate this adoption? Is there a “killer app” for this technology that will trigger a sudden increase in deployments?