Coping with link failures in centralized control plane architectures

Abstract

Recently proposed SoftRouter and 4D network architectures recommend having the least amount of intelligence in the switches. They advocate transferring control plane functionalities to general-purpose servers that could govern these switches remotely. A link failure in such architectures could result into switches losing contact with the controller or even generating routing loops. These scenarios could generate a large amount of unnecessary traffic in the network. We study the implications of a failed link in such architectures. We develop an algorithm that would inform only the relevant switches to refrain from sending messages in the direction of the failed link, and yet have the minimum amount of intelligence on the switches. We implement our algorithm on a network formed by OpenFlow switches and evaluate its performance. Our experiments verify that the performance of our algorithm is dependent on the total number of flow-table entries in a switch. We also verify that by implementing our algorithm all the necessary switches are informed of the failed link significantly sooner than the controller identifies the failed link and sends out an update.