— We consider the general problem of optimizing the performance of OBS networks with multiple traffic classes subject to strict (absolute) QoS constraints in terms of the end-toend burst loss rate of each guaranteed class of traffic. We employ Markov decision process (MDP) theory to obtain optimal wavelength sharing policies for two performance objectives, namely, maximization of weighted network throughput and minimization of the loss rate of best-effort traffic, while meeting the QoS guarantees. The randomized threshold policies we obtain are simple to implement and operate, and make effective use of statistical multiplexing. In particular, the threshold randomization feature enables the policies to allocate bandwidth at arbitrarily fine sub-wavelength granularity, hence making effective use of the available network capacity.
Li Yang, George N. Rouskas