The scalability issue with the #SDN model today isn’t that the data plane that won’t scale …it’s issues with the control plane.

Reading “OpenFlow/SDN Is Not A Silver Bullet For Network Scalability” brought to light an important note that must be made regarding scalability and networks, especially when we start talking about the control plane. It isn’t that the network itself won’t scale well with SDN, the concern is – or should be – on the control side, on whether or not the integration of the control plane with the data plane will scale.

sdntopology

A core characteristic of SDN is not only the separation of the control and data planes, but that that the control plane is centralized. There can be only one. The third characteristic that is important to SDN is the integration of these decoupled data plane devices with the control plane via APIs (Mike Fratto does an excellent job of discussing the importance of API support as well as making the very important distinction between API and SDK in his recent blog, “Three Signs of SDN Support to Watch for from Vendors”, so I won’t belabor this point right now).

The convergence of these three characteristics results in what Enterprise Application Integration (EAI) has long known as a “hub and spoke” integration pattern. A hub – in the case of SDN, the controller – sits in the middle of a set of systems – in the case of SDN, the data plane devices – and is the center of the universe.

The problem with this pattern, and why bus topologies rose to take its place, is that it doesn’t scale well. There is always only one central node, and it must necessarily manage and communicate with every other node in the integration. While hub-and-spoke, which grows linearly, isn’t nearly as difficult to scale as its predecessor the spaghetti (mesh) pattern, which grows exponentially, in a network growing even linearly is going to be problematic for some value of n (where n is the number of edges, i.e. nodes, in the network).

At some value of n, it becomes apparent that the controller (hub) must be able to scale, too. Scaling up would require expansion of the system upon which the controller is deployed, which may require replacement. You can imagine the reluctance of operations to essentially shut down the entire network while that occurs. The other option is to scale out, vis à vis traditional methods of scaling other systems, via a load balancing service and duplicate instances. This implies a shared-something architecture, usually describes as being the database or repository for policy from which nodes are “programmed” by the controller. This appears to be the response in existing implementations, with “clusters of controllers” providing the scale and resiliency required.

So scaling out the controller, then becomes an exercise in traditional scalability methods used to scale out client-server architectures.

So the Control Plane of an SDN Can Scale. What’s the Problem then?

As pointed out by Ivan Pepelnjak in “OpenFlow/SDN Is Not A Silver Bullet For Network Scalability”, the problem with this model appears to be response time. Failure in a node cannot be addressed fast enough by a centralized software system, particularly not one that relies on a database (which has its own scalability issues, of course).

There are several questions that must be answered in order to even deal with failure that pose some interesting performance and scaling challenges.

  1. How does the controller know that an element node is “down”? Is it polling, which introduces an entirely new concern regarding the level of monitoring noise on the network interfering with business-related traffic? Is it monitoring a persistent control-channel connection between the controller and the node? Certainly this would indicate nearly instantaneously the status of the node, but introduces scaling challenges as maintaining even a one-to-one control-channel connection per element node in the network would consume large quantities of memory (and ultimately have a negative impact on performance, requiring scale out much sooner than may be otherwise necessary). Does a neighbor or upstream element  tattle on the downed node when it doesn’t respond?

    There are a variety of mechanisms that could be used to monitor the network such that the controller is informed of a failure, but each brings with it new challenges and has different responsiveness profiles. Polling is tricky, as any load balancing provider will tell you, because it’s based on a timed interval. Persistent connections, as noted earlier, bring scalability challenges back to the table. Tattle-tale methodologies are unreliable, requiring that a neighbor or upstream element have the need to “talk to” the downed down before notification can occur, leaving open the possibility of a downed node going unnoticed until it’s too late.
  2. How does the controller respond to a downed element node? Obviously the controller needs to “route around” or “detour” traffic until a replacement can be deployed (virtually or physically). This no doubt requires some calculations to determine the best route (OSPF anyone?) if done in real-time. Some have suggested alternative routes in tables be available on each node in the event of a failure, a model more closely related to today’s existing routing practices and one that would certainly respond much better to failure in the network than would a system in which the controller must discover and reconfigure the network to adjust to failures.
  3. What happens to existing flows when an element node fails? Ah, the age old stateful failure challenge. This is one that is (almost) solved with redundant architectures that mirror sessions (flows) to a secondary device. The problem is that these models work best, i.e. have the highest levels of success, for full-proxy devices, particularly when the flow supports stateful/connection-oriented protocols. 

These questions are nothing new to experienced EAI practitioners who’ve had to suffer through a hub-and-spoke based integration effort. Failure in a node or of the controller give rise to painful fire-drill exercises, the likes of which no one really enjoys because they are highly disruptive. They’re also not really new questions for those with a long history in load balancing and high availability architectures.

Still, these are questions which need to be answered in the context of the network, which has somewhat different uptime and performance requirements than even applications.

Ultimately the answer is going to lie in architecture, and it’s unlikely that what results will be a single, centrally controlled one.