Tech giants are pitching in to fund a research center for software-defined networking. Google just revealed it took the OpenFlow plunge for its internal networks starting in 2010.
The idea of software-defined networks (SDN) began six years ago as a research initiative by two professors from Stanford and UC Berkeley. A year ago a number of industry giants joined in kicking off the Open Networking Foundation to develop and promote OpenFlow, the open-source software instantiating the academics' ideas. Last week a partially overlapping group of networking heavyweights announced backing for the Open Networking Research Center, based at Stanford and Berkeley, to create the intellectual foundation of SDN. And this week the ONF held a technical conference in Santa Clara, CA, at which Google took the wraps off of a 2-year infrastructure project, now nearly complete, to convert its entire datacenter WAN to OpenFlow control.
Google is usually mum about its operations-level details, so it was unusual for the company's infrastructure czar Urs Hölzle to speak in public at all. According to Wired, Hölzle said that the idea behind OpenFlow is "the most significant change in networking in the entire lifetime of Google." In the course of the talk, the Google SVP revealed that the company has been designing and building its own networking hardware for the project, because no one was selling "dumb" hardware that met its requirements.
What is software-defined networking? The idea is to dumb down the routers in a network so that they do only switching; control of flows, and of the network as a whole, is then implemented as software that presents an API for use by operators or applications. Or as Hölzle told Wired:
You have all those multiple devices on a network but you're not really interested in the devices -- you're interested in the fabric, and the functions the network performs for you. Now we don't have to worry about those devices -- we manage the network as an overall thing. The network just sort of understands.
Commodity hardware and increased visibility and control of network flows sound like a great deal for the network builder. It does not sound so great to established networking vendors, whose profit margins and ability to lock in customers reside in their proprietary hardware.
Cisco has maintained a distance from the OpenFlow action, though the company supports both the ONF and the ONRC. Recently the New York Times reported that Cisco is incubating a "spin-in" called Inseimi to introduce OpenFlow into its product lines. Serial entrepreneur William R. Koss last month blogged about his speculation on two possible strategies Cisco could be pursuing with Insiemi. In one of them, he imagines Inseime building "session border control" between an OpenFlow network and Cisco's proprietary Nexus OS-controlled network. With such a product Cisco could keep existing customers locked in while promising access to the flexibility of OpenFlow.
From another legacy networking supplier, here is Markus Nispel, chief technology strategist at Enterasys Networks (a Siemens Enterprise Communications company), quoted in NetworkWorld.com:
The main concern here is trying to externalize the flow setup. 10G Ethernet could bring you up to 15 million flows per second per interface. How could an external system cope with that? We do have hardware assistance internally to the system to manage flows in the system. External [would be] challenging.
Google does not seem to be having any such issues with OpenFlow, according to a NetworkWorld.com report of Urs Hölzle's talk at the Open Networking Summit this week. The company's datacenter WAN is huge -- if it were an ISP it would be the second-largest in the world, behind Level 3, with its 100,000 miles of fiber. Hölzle reported that the transition to OpenFlow occurred with unbelievable smoothness, largely because OpenFlow gave them the ability easily to simulate backbone-scale network environments virtually, along with the ability to mirror production event streams into testing environments. Hölzle's final slide declared flatly that "OpenFlow is ready for real-world use." The Google SVP said that the benefits of OpenFlow so far include improved manageability and improved costs, though it is too early to give exact figures.
For an enterprise to transition a data center now to OpenFlow would be an adventurous move that could pay off in a substantial agility boost. Google claims that it is getting close to 100 percent utilization on its datacenter WAN. This is unheard-of: most networks are designed to operate at 30 or 40 percent in order to build in overhead for bursty traffic. Google also claims that the network gear it designed is up to 50 times as fast as what is available commercially.
Few OpenFlow products are on the market today (among them are some jointly marketed by NEC and IBM, sponsor of this site), but many more are expected by year's end. Cloud providers are likely to be among the first to dive into OpenFlow. Software-defined networks are coming soon to a supplier near you.