A Spare Capacity Planning Methodology for Wide Area Survivable Networks

Abstract

In this dissertation, a new spare capacity planning methodology is proposed utilizing path restoration. The approach is based on forcing working flows/traffic which are on paths that are disjoint to share spare backup capacity. The algorithm for determining the spare capacity assignment is based on genetic algorithms and is capable of incorporating non-linear variables such as non-linear cost function and QoS variables into the objective and constraints. The proposed methodology applies to a wider range of fault scenarios than most of the current literature. It can tolerate link-failures, node-failures, and link-and-node failures. It consists of two stages: the first stage generates a set of network topologies that maximize the sharing between backup paths by forcing them to use a subset of the original network. The second stage utilizes a genetic algorithm to optimize the set of solutions generated by the first stage to achieve an even better final solution. It can optimize the solution based on either minimizing spare capacity or minimizing the total network cost. In addition, it can incorporate QoS variables in both the objective and constraints to design a survivable network that satisfies QoS constraints.

Numerical results comparing the proposed methodology to Integer Programming techniques and heuristics from the literature are presented showing the advantages of the technique. The proposed methodology was applied on 4 different size networks based on spare capacity optimization criteria and it was found that it achieved solutions that were on average 9.3% better than the optimal solution of the IP design that is based on link-restoration. It also achieved solutions that were on average 22.2 % better than the previous heuristic SLPA.

The proposed methodology is very scalable. It was applied on networks with different sizes ranging from a 13-node network to a 70-node network. It was able to solve the 70-node network in less than one hour on a Pentium II PC. The curve-fitting of the empirical execution time of the methodology was found to be O(n3).