Maintaining an Up-to-date Global Network View in SDN

TL;DR

This paper investigates how network state collection and distribution affect SDN controller performance, highlighting that outdated information and increased controller count degrade application effectiveness.

Contribution

It provides a comparative analysis of active and passive state collection methods and their impact on SDN load-balancing applications through modeling and real-world evaluation.

Findings

Both state collection and distribution impact SDN application performance.

Outdated network information leads to higher performance degradation.

Distributed controllers amplify the effects of state collection methods.

Abstract

Maintaining an up-to-date global network view is of crucial importance for intelligent SDN applications that need to act autonomously. In this paper, we focus on two key factors that can affect the controllers' global network view and subsequently impact the application performance. Particularly we examine: (1) network state collection, and (2) network state distribution. First, we compare the impact of active and passive OpenFlow network state collection methods on an SDN load-balancing application running at the controller using key performance indicators that we define. We do this comparison through: (i) a simulation of a mathematical model we derive for the SDN load-balancer, and (ii) an evaluation of a load-balancing application running on top of single and distributed controllers. Further, we investigate the impact of network state collection on a state-distribution-aware load-balancing application. Finally, we study the impact of network scale on applications requiring an up-to-date global network view in the presence of the aforementioned key factors. Our results show that both the network state collection and network state distribution can have an impact on the SDN application performance. The more the information at the controllers becomes outdated, the higher the impact would be. Even with those applications that were designed to mitigate the issues of controller state distribution, their performance was affected by the network state collection. Lastly, the results suggest that the impact of network state collection on application performance becomes more apparent as the number of distributed controllers increases.