Neural scheduling algorithms for time-multiplex switches

In an N×N time-multiplex switch, transmission conflict arises when two or more input adaptors transmit packets to the same output adaptor simultaneously. To resolve transmission conflict, we propose two neural-based scheduling algorithms which use a large number of simple processing elements to perform scheduling in parallel. The first algorithm uses N² hysteresis McCulloch-Pitts (1943) neurons to determine conflict-free transmission schedules with maximum throughput. The second algorithm resolves transmission conflict among the first M packets in each input queue. It determines suboptimal transmission schedules using only NM neurons (M<N). M is a design parameter: if M is larger, we can find closer-to-optimal transmission schedules, but we need more neurons. Simulation results show that the first algorithm can find near-global optimal transmission schedules. The second algorithm can give close-to-optimal transmission schedules using only a small M. When N=500 and M=10, the throughput efficiency is already 96.44% while the required number of neurons is reduced from N ²=250000 to NM=5000