SDN and the Virtual Network Instance (VNI)

Software Defined Networking (SDN) is a relatively new term and its definition is still being refined. I thought that it would be useful to help refine the definition with some examples of SDN as I see it.

My thinking on SDN was refined by watching Scott Shenker’s presentation “A Gentle Introduction to SDN”: http://www.youtube.com/watch?v=eXsCQdshMr4

Scott makes the point that the networking industry needs new abstractions that hide the complexity of networking. Good abstractions hide the complexity of the underlying systems, much like the seven layer networking model where each layer hides the complexity of the lower layers. The best abstractions help us construct simple mental models that allow us to work much more efficiently. The result is enormous productivity gains. Good examples are the gains that were achieved by switching to virtual memory or the benefits from using virtual machines for compute processing.

I’ve started to think of SDN as an abstraction system that allows me to think of the network virtually. In other words, I can create a Virtual Network Instance (VNI) and treat it as if it were a real network (see Sidebar) without regard to the complexity of the underlying physical network.

Sidebar: A definition of ‘virtual’: Something that you can treat as if it were real, but is really an abstraction of something else. Example: Virtual Memory is an abstraction of a large memory system which is comprised of a variety of storage mechanisms such as RAM and Disk. Some efficiency is generally lost in the abstraction’s implementation, but the efficiency of using the abstraction more than offsets the implementation inefficiency.

While watching Scott Shenker’s presentation, the power of the abstractions that SDN can provide became obvious. I could see the parallels between a VNI and other virtual abstractions. It would be great to be able to hide the complexity of creating a Layer 3 network, which typically requires creating a VLAN, assigning an IP subnet, configuring redundant routers, updating the routing protocols, and perhaps using MPLS to tie it to other subnets in a virtual Layer 3 domain. Why can’t a system be built that automates the configuration and hides the complexity?

In traditional network configurations, the network design is frequently dictated by how it is to be used. For example, many data centers are being designed as a large Layer 2 domain because of the requirement for easy migration of Virtual Machines (VMs) to any place within the data center. We know that large L2 domains have problems, yet we are driven by the VM migration requirement to implement them. This is silly and counter-productive. It is much better to create scalable, stable designs that utilize networking strengths and avoid weaknesses.

We should be designing networks that limit failure domains. The protocols we use should work well and we should avoid those protocols that don’t work well or that don’t scale to fit our needs. Redundancy can be built into the physical network where we know it is needed. The physical topology should provide the necessary bandwidth with fast failover where needed. QoS and policy routing should allow multiple VNIs to operate in parallel, with each specifying its own operational requirements. Design the physical network for resiliency so that when a failure occurs, the VNIs continue to operate. The complexity of the underlying physical network is then hidden by the VNI implementation.

I foresee a time when VNI definitions could be created with object-oriented programming languages such as Python, Ruby, or OO-Perl. These languages would then connect to the underlying SDN automation engines (e.g., an OpenFlow NorthBound API). The set of switch ports that are to be in a given network could be specified as one object, then the VNI object could be defined to provide the desired connectivity of those ports. Getting these definitions right will take some time, just as it took time to get virtual memory and VM definitions right.

-Terry