A Neural-network Controller for Scheduling Packet Transmissions in a Crossbar Switch

A Hoppeld neural network is presented for scheduling packets in an input-buuered crossbar switch. The controller has a two-dimensional winner-take-all design and is driven by request signals representing the packets stored in the input buuers. Schemes are presented to support multiple priorities of traac by changing the binary request signals to multi-valued signals that vary over a limited range. Simulation results show that the controller provides close-to-ideal through-put for switches as large as 32 32. In addition, by varying the relative strength of request signals, the scheme is shown to be able to control jitter eeectively in a communication network consisting of multiple switches in series.