Rethinking the Network

Marten Terpstra

Subscribe to Marten Terpstra: eMailAlertsEmail Alerts
Get Marten Terpstra: homepageHomepage mobileMobile rssRSS facebookFacebook twitterTwitter linkedinLinkedIn


Related Topics: Virtualization Magazine, SDN Journal

Article

Resiliency in Controller-Based Network Architectures

At the core of SDN solutions is the concept of a controller

Last week Ivan Pepelnjak wrote an article about the failure domains of controller based network architectures. At the core of SDN solutions is the concept of a controller, which in most cases lives outside the network devices themselves. A controller as a central entity controlling the network (hence its name) provides very significant values and capabilities to the network. We have talked about these in this blog many times.

Centralized Control

When introducing a centralized entity into any inherently distributed system, the architecture of such a system needs to carefully consider failure domains and scenarios. Networks have been distributed entities, with each device more or less independent and a huge suite of protocols defined to manage the distributed state between all of them. When you think about it, it’s actually quite impressive to think about the extend of distribution we have created in networks. We have created an extremely large distributed system with local decision making and control. I am not sure there are too many other examples of complex distributed systems that truly run without some form of central authority.

It is exactly that last point that we networking folks tend to forget or ignore. Many control systems in the world have central control and management. And the vast majority of them work pretty well. Any complex manufacturing facility has centralized control over robots, belts and all other machinery that it may use. There usually is some distributed state and health checks at interfaces between machines and operations, but the entire end to end process is controller by a centralized entity.

The reason for this is not much different from the reason we are starting to deploy controllers in networks. Having a true end to end view of all available resources will provide better overall performance of and control over the network. A centralized entity can make choices and decisions that are related or dependent of previous choices based on information that may well be outside the reach of a typical system in distributed operation.

Architectural Choices

But the introduction of such an entity needs to be carefully architected and designed. The exact role of a controller in the day to day (or microsecond to microsecond) operation of a network becomes a critical choice, it defines the dependency of the network on the controller and as a result, the impact of a controller failure. At Plexxi we made a very deliberate architectural choice for our controller:

  • it cannot ever be in the data path of network traffic. Not for new flows, not for existing flows. Not for link failures. Not for switch failures.

The network has to run when the controller is not available. It has to run for existing attached devices, newly attached devices, existing flows and new flows. Of course we want the controller to be available all the time because it gives us the best visibility, but we very deliberately architected it so that the network keeps working if it isn’t.

To that purpose we split our controller into two separate components. The most visible (and perhaps even traditional in this new world of controller architecture) is our central controller. It’s software, runs on a VM or bare metal server and is the central coordinator. It maintains the database with all relevant data. It communicates with the switches. And the operator communicates with it through a GUI, or our APIs or Data Services Engine.

Then there is a distributed portion of the controller. It run on every Plexxi Switch. It communicates with the central controller and takes higher level configuration, policy and topology instructions, then passes them to the Switch software that turns this into configuration for the hardware etc. Similarly, things like statistics and state info from the Switch software is passed to the distributed portion of the controller, then passed back to the central controller.

Network Independence

But most importantly, Plexxi switches are fully capable of making forwarding decisions by themselves. They learn MAC addresses. They resolve ARP. They have L2 forwarding tables. They have L3 forwarding tables. And these tables themselves are not managed by the central controller. They are managed by each switch. What the central controller provides is topology information on how to reach other switches in a Plexxi domain. Out of the many paths through the fabric, which ones should be used and for what percentage of traffic. And hundreds of backup paths through that fabric if a link of switch fails. And those failures are communicated between the switches themselves, without involving the controller (who gets informed, but is not in the action path).

Having this very clear line in the sand of what the switches are responsible for and what the controller is responsible for allows us to worry (just a little) less about the 100% resiliency of the controller. Don’t get me wrong, we want the controller there, but your network will operate as you expect if its not. In his article, Ivan calls it “controller enhanced network infrastructure”. That works.

[Today's Fun Fact: All polar bears are left-handed. Or left-clawed. I would assume that means they tend to be more creative than other bears too.]

The post Resiliency in Controller based Network Architectures appeared first on Plexxi.

More Stories By Marten Terpstra

Marten Terpstra is a Product Management Director at Plexxi Inc. Marten has extensive knowledge of the architecture, design, deployment and management of enterprise and carrier networks.