Information theoretic approach to traffic adaptive WDM networks

WDM networks adapt to the changes in traffic by reconfiguring the virtual topology. Though reconfiguration is done with the objective of utilizing resources efficiently, the resulting disruption in traffic is a cause for concern. Hence, policies are formed to decide on the time (i.e., when) to trigger reconfiguration and the new virtual topology that is most beneficial to the network. We present a simple, general and flexible framework, based on the two conflicting objectives of efficient resource utilization and minimizing traffic disruption, to evaluate reconfiguration policies. Instead of re-determining the reconfiguration policy whenever the traffic changes, we present Incremental Clustering Algorithm (ICA) to pre-plan the reconfiguration policy for a fully predictable finite sequence of traffic matrices. Since full predictability of such a sequence is not possible in practice, we learn the traffic sequences in order to probabilistically predict the future ones. From an information theoretic point of view, we quantify the predictability of traffic sequences and the number of times the reconfiguration policy is re-determined for any WDM network. To optimally predict the future traffic sequences and to incur optimal cost in the re-determination of the reconfiguration policy, we propose Universal Reconfiguration Management System (URMS). A Prediction-based Incremental Clustering Algorithm (PICA) that extends ICA is used by URMS to predict the reconfiguration policy. Within URMS, the probabilities are assigned to the traffic sequences by the prediction schemes of LZ78. We performed extensive simulations to study the effectiveness and efficiency of URMS when compared to the fully predictable and totally unpredictable models. The performance of URMS improves with learning and nearly achieves the performance of a fully predictable model.