Simulations of Crossbar Switches for Parallel Systems

With asynchronous packet arrivals, we nd that, when the number of packet its is close to the fan-in of the crossbar switch, performance of a non-multiport central bu er switch becomes comparable to its multiport counterpart. Assuming the same amount of bu er memory, simulations of oblivious routing on hypercube indicate that split input bu ering outperforms central bu ering, even when dedicated bu ers for each (source, destination) pair are provided. This e ect, common in other topologies as well, is caused by the low utilization of bu er memory for central bu ering.