Performance of a neural net scheduler used in packet switching interconnection networks

An analysis of an interconnection network for packet switching controlled by a neural network is presented. The interconnection network is a crossbar switch of size N which maintains N queues at each of the inputs. The neural net controller is of the Hopfield type, and performance evaluation is done under realistic input traffic assumptions. In order to investigate the influence of the random characteristics of the packet traffic on the neural net performance, the offered packet streams are modeled by means of batch inputs, where the inter-arrival time and the packet length can be arbitrarily chosen. The incoming traffic indicating source-destination packet stream is symmetrically distributed over the crosspoints of the interconnection network. Simulation results for various system parameters are presented with respect to the performance measures, such as switch throughput and transfer delay. The performance of the neural net as switch fabric controller is compared to conventional control schemes, where the issue of scheduling fairness is addressed.<<ETX>>