Written by Dan Pitt
Dan Pitt is Executive Director of the Open Networking Foundation, joining on its public launch in March 2011.
In this segment, Dan Pitt delves into the SDN model and describes where decisions are made within it.
The ultimate decisions will be made really at the interface or just below the interface between the corporate applications and the control algorithms for the network. What we expect these interfaces to look like - and there's not going to be a single one, we called them northbound interfaces - will be vehicles for conveying the network's capabilities and state up to the applications, and the application requirements down to the network. We don't want the applications to have to know the specifics of the network. We've abstracted it expressly to avoid that. We don't want the network to know the details of the application, just what it expects from the network. I think as we have a better library of generic network behaviors and generic application requirements we can then very easily develop optimization algorithms in the control plane that make sure that the behaviors that it tells the network to exhibit are the ones that fulfill the objectives of the applications.
Where we have an interest is in maintaining the notion that the intelligence of the model you might have of the network is located in the control plane and not in the forwarding plane, because we want the forwarding plane to be simple. We'd rather have the model of the network convert the application requirements to network behaviors in a consistent way in the control plane and the behaviors just conveyed to the forwarding plane, rather than have the forwarding plane have to interpret the model and convert that to behaviors. That's more intelligence than we think we want, is ideal to have in the forwarding plane. Forwarding plane needs to be able to execute packet processing according to instructions it receives and to be able to measure what's happening in the network in important ways and convey that up to the control plane.
I don't think operators should worry too much about which model to adopt. I think as we get into more software and open source software those things can change and iterate quite rapidly. We will learn in fairly short order what works and what doesn't. I'm hoping that we will capture a lot of this knowledge in a series of libraries, library files that are open source that will essentially house the information models and in the software the data models for certain uses across certain interfaces, certain application requirements and certain network behaviors. Then you can pick and choose from these software modules what to instantiate when you build a network control plane.
Interesting about OpenFlow: it's the southbound protocol between the separated control plane and forwarding plane. It's like the drive shaft of a car. You won't move if the engine and rear wheels aren't connected, but the drive shaft is not the exciting part of the car. OpenFlow is really necessary but the exciting part of SDN is how you map it to organizational objectives.