Presenter: Dimitri Papadimitriou
The control model of networks such as the Internet can be explained as a static open-loop control. While such model has been effectively used since so far, this model is progressively reaching its limits. Indeed, once configured, routers follow explicitly pre-defined behavior, persistently decide and uniformly execute, leaving no possibility to locally diagnose their internal state, behavior, and environment over time so as to adapt their decisions and executions. This mode of operation is particularly detrimental during the convergence of the routing states when simultaneous failures occur or during the update of the routing table entries resulting from a link or node failure. In this talk, we present the performance evaluation results obtained when applying on-line learning in combination with adaptive resilience techniques. Our results show that this approach enables to anticipate the detrimental effects of such events so as to improve network availability.