On placement of controllers and hypervisors in software defined networks
No Thumbnail Available
Software defined networking shifts the control plane of forwarding devices to one or more external entities known as controllers. The placement of controllers in the network influences every aspect of a decoupled control plane, from state distribution options to fault tolerance to performance metrics. Determining the number and placement of controllers is an important problem in software defined networking. Failure of a controller results in disconnections between the controller and the switches that are assigned to it. The administrator can reassign each switch of the failed controller to a working controller with enough capacity that is nearest to the switch. However, the reassignment of switches result in a significant upsurge in the worst case latency.In this thesis, we propose optimization models for the failure foresight capacitated controller placement that avoids disconnections, repeated administrative intervention, and drastic increase in the worst case latency in case of controller failures by maintaining a list of μ(> 1) reference controllers for every switch. The objective is to minimize the worst-case latency between switches and their μth reference controllers while satisfying the capacity and closest assignment constraints. First, we design an optimization model for a single controller failure and extend it to multiple controller failures. We also design a variant of failure foresight capacitated controller placement that minimizes the sum of worst-case latencies from switches to their 1 st , 2 nd ,. . . , μth reference controllers. Next, we relax the failure foresight assumption of switches and investigate a capacitated next controller placement strategy that not only considers capacity and reliability of controllers but also plans ahead for controller failures. We design an optimization model for a single controller failure and extend it to multiple controller failures. We also present a simulated annealing heuristic to produce fast and viable solution on large networks.
Supervisor: S. V. Rao
COMPUTER SCIENCE AND ENGINEERING