< back to overview

Single Point of Failure. Not.

Apr 29, 2014
Sue Kim - gu
Sue Kim - gu About the author

Luxoft’s Ash Bhalgat shares his views on the SDN controller and single point of failure concerns. 

A central control plane in the OpenFlow® based SDN architecture is sometimes described (inaccurately) as a single point of failure.  Recently, some of the industry players have proposed SDN architectures that would split network control functionality between data plane and control plane devices. The argument is that this would mitigate the risk of single point of failure. Although a split control plane is a fascinating idea (and probably fits some use cases), it would not achieve true high availability network architecture. Further, such split control plane architecture wouldn’t offer benefits of a centralized control plane such as broader visibility of network resources and agile programmability that would result in lower cost of operations. There are ways to mitigate single point of failure risk by designing a highly available (HA) control plane and we have seen such systems since TDM days.

ONF Executive Director Dan Pitt recently mentioned a distributed computing example in the ONF blog about a paradigm shift from de-centralized to centralized control plane. I would like to further share an example of similar paradigm shift that happened in the networking industry itself.  The Wireless LAN (WLAN) industry went through a major transition from thick or autonomous or heavyweight Access Points to thin or lightweight access points (LAP) around 2005. Before 2005, the WLAN industry primarily practiced de-centralized WLAN architecture with autonomous access points managing both Layer 1 (radio frequency) and Layer 2 (media access control) functions. Clearly, this architecture couldn’t scale and customers faced challenges such as seamless radio resources management (RRM) and consistent policy enforcement. Some of the policy enforcement inconsistencies even resulted in security breaches. Essentially, the WLAN industry needed a better approach to control, manage, and scale enterprise and service provider WiFi networks and a departure from the de-centralized model was a need of time in 2005. A better approach would also help to drive WiFi technology adoption from niche to major markets, which it did.

The Wireless LAN Controller or WiFi Switch emerged as an elixir to the many shortcomings of the autonomous access point-based de-centralized WLAN architecture. Essentially, a WLAN controller centralized the RRM enforcement functions offering a single point of control for enterprises and service providers. Moving network intelligence from heavy data plane (autonomous access points) to dedicated control plane (WLAN controller) resulted in thinner data plane (LAP), lowering costs and making the network more responsive to operator needs. To embrace this technology transition, industry players made strategic acquisitions: Cisco acquired Airespace in 2005, HP acquired Colubris in 2008, and Juniper acquired Trapeze in 2010. WLAN industry players even created management tools to migrate from older de-centralized architecture to newer centralized architecture and facilitate faster adoption to controller based WLANs. As a result of this paradigm shift, the WLAN market size doubled from $1.5B in 2005 to $3B in 2008. All mission-critical WiFi networks today are running a centralized model, and these WiFi networks haven't failed their customers. While there have been security vulnerabilities in these WiFi networks due to protocol shortcoming that led to credit card data thefts, networks have been operational at all times.

Today, the overall networking industry is facing similar challenges that plagued the WLAN industry in 2005. We need networks that are flexible, programmable, and manageable. Open SDN architecture is a step in that direction. The benefits of a centralized control plane far outweigh the risk of single point of failure, which in itself can be mitigated with high availability systems, as nearly every commonly used online service demonstrates.

In summary, I would like to emphasize that centralization is not really a new concept but just a need of time as underlying networks evolve to keep up with application demands.

- Ash Bhalgat, Senior Director, SDN Products, Luxoft

Share this post:
Sue Kim - gu