Network Virtualization - Then and Now ... and Future

Recent events in Software Defined Networking (SDN) - not the least of which was major vendors including Cisco and Juniper partner with universities to launch the Open Networking Research Center (ONRC) - have spurred me to take a closer look at the hype.

It's not like I was totally in the dark about this. "Network Virtualization" has been talked about since just after server virtualization hit the scene. It was more of a theory or pie-in-the-sky idea back then - no one really could articulate the real benefits in terms of cost or new functionality, but if you virtualized servers, networks were the next logical step - right?

So today, we hear about OpenFlow, which is just one implementation of network virtualization. How does all this fit in? What is the evolution? This is the way I break it down:

• Network Virtualization
• Software Defined Networking
• OpenFlow
• ...
• Multi-Topology Routing
• ...

So what's this multi-topology routing? Apparently, it's been around since at least 2007 and ready to go on Cisco devices, among other vendors. In fact, when I was working with mobile ad-hoc networks (MANET) from 2009 to 2011, we looked at various new routing protocols and enhancements to existing routing protocols to compensate for the low bandwidth, unreliable links typical in a MANET. One option was multi-topology versions of IS-IS and OSPF which would create logically separate routing tables over the same physical network links to provide routing based on user-defined categories. For example, one routing table would use higher bandwidth links and route only video traffic while another would use low bandwidth but more reliable links for voice. This would happen simultaneously over the same physical network.

The main difference from this network virtualization approach from the SDN approach that OpenFlow is championing is the control plane. In the multi-topology routing example, the routing processes are still distributed between the nodes that create the network. The network nodes need to be "intelligent"; they need to run a routing process, communicate with neighbors, maintain adjacencies and create a localized view of the network. With SDN and OpenFlow, the control plane is removed from the distributed node model and centralized in a controller.

This has some advantages. From experience, configuring multi-topology routing is complex and the more nodes in the network increases the difficulty and the chance for errors. A centralized controller that automatically distributes routing policy to small footprint agents on commodity network nodes eliminates the configuration distribution issue while also reducing costs of the network deployment. This starts to answer the "benefits in terms of cost or new functionality" question earlier in this post.

Additionally, with treating streams of similar packets as "flows", we can realize the benefits of a circuit switched model where routes through a network can be defined not just by the least hop count, but by other factors such as reliability of links, actual load on links and delay sensitive issues for real-time traffic such as voice and video. This all without compromising or changing the packet nature of IP(v6) networks.

Great - so where do I sign up? Not so fast. While research in this area is hot - trade magazines and conferences are highlighting OpenFlow and major vendors like Cisco and Juniper are taking notice, I don't expect a revolution soon. Places you'll see this first are certainly academia and high-tech content hosting companies with large-scale complex data centers. As always, enterprise adoption will be halted until a major vendor has an offering. And don't confuse Cisco's and Juniper's support of the ONRC as an immanent product offering. SDN has some real implications for major vendors. The removal of the routing smarts to a centralized controller means 'dumb' switches will suffice and for companies that make money off intelligent switches, a migration to SDN will hurt their bottom line unless they can control their destiny and entrance into the SDN fray.

And there are other mitigating factors that will push SDN in the enterprise off for about 5 years. The outsourcing of corporate networks to providers means one place SDN makes sense - the WAN - is offloaded. The other place - the data center - is another outsourcing play between all forms of cloud computing. Enterprises that maintain their own data centers are just coming to grips with data center "fabrics". Typical capital expenditures for networking components are closer to 5 years. Besides, those heavy into virtualization may already be using "network virtualization" - the Cisco Nexus 1000V.