Abstract
Motivated by the need for a simple switch architecture that scale with the speed of fiber optics, we consider two switcharchitectures that use both input expansion and output expansion. The first one, called the (K1,K2)-switch in this thesis, expands every input port into K1 queues and every output port into K2 queues. For this, a K1N x K2N switching fabric is needed, where N is the number of input/output ports. The second one, called the (k1,K2)-switch, only expand every input port into k1 queues, and it only needs an N x K2N switching fabric. Via extensive simulations, we compare their performance, including throughput and average delay, under various traffic models. From these simulations, we conclude that both the (k1,K2)-switch and the (K1,K2)-switch with k1 >= 2, K1>=2, and K2>=2 perform very well if the switch only supports unicast traffic. The average delays for such switches are almost identical to that of the output-buffered switch as the bottlenecks for such switches are the output links. For non-uniform multicast traffic, we find that the (K1,K2)-switch is substantially better than the (k1,K2)-switch with k1 =K1. This is due to the fact that the queue implementation in the(k1,K2)-switch cannot handle a large number of fanouts at theinput ports.