Future Internet (Mar 2019)

Dynamic SDN Controller Load Balancing

  • Hadar Sufiev,
  • Yoram Haddad,
  • Leonid Barenboim,
  • José Soler

DOI
https://doi.org/10.3390/fi11030075
Journal volume & issue
Vol. 11, no. 3
p. 75

Abstract

Read online

The software defined networking (SDN) paradigm separates the control plane from the data plane, where an SDN controller receives requests from its connected switches and manages the operation of the switches under its control. Reassignments between switches and their controllers are performed dynamically, in order to balance the load over SDN controllers. In order to perform load balancing, most dynamic assignment solutions use a central element to gather information requests for reassignment of switches. Increasing the number of controllers causes a scalability problem, when one super controller is used for all controllers and gathers information from all switches. In a large network, the distances between the controllers is sometimes a constraint for assigning them switches. In this paper, a new approach is presented to solve the well-known load balancing problem in the SDN control plane. This approach implies less load on the central element and meeting the maximum distance constraint allowed between controllers. An architecture with two levels of load balancing is defined. At the top level, the main component called Super Controller, arranges the controllers in clusters, so that there is a balance between the loads of the clusters. At the bottom level, in each cluster there is a dedicated controller called Master Controller, which performs a reassignment of the switches in order to balance the loads between the controllers. We provide a two-phase algorithm, called Dynamic Controllers Clustering algorithm, for the top level of load balancing operation. The load balancing operation takes place at regular intervals. The length of the cycle in which the operation is performed can be shorter, since the top-level operation can run independently of the bottom level operation. Shortening cycle time allows for more accurate results of load balancing. Theoretical analysis demonstrates that our algorithm provides a near-optimal solution. Simulation results show that our dynamic clustering improves fixed clustering by a multiplicative factor of 5.

Keywords