A SUBOPTIMAL STRAIGHT MATCHING METHOD

CLOS-NETWORK SWITCHES 258

10.2 A SUBOPTIMAL STRAIGHT MATCHING METHOD

FOR DYNAMIC ROUTING A crucial requirement on the routing or scheduling algorithm is that it must run extremely fast, because cells have to be assigned paths on a slot-by-slot basis. A distributed routing algorithm is considered, which is suboptimal in performance, but runs fast and is simple to implement in hardware. Let us consider a particular input module, say the first one. It may have a total of n cells arriving at a time, and the assignment hardware has to finish assigning these n cells within some prescribed duration ⌬ less than a time slot. We further divide ⌬ into k mini-slots of length ⌬rk, and now focus on the assignment operation inside a minislot. During every mini-slot, the B of j a specific output module is made available for assignment to an input module A , i.e., matching of A with B . Note that there are at most n matches i i j needed in each mini-slot, as there are at most n input cells at an input module. While A is matched against B , simultaneously A is matched i j iq1mod k against B , and so on, thereby permitting parallel assignments. In the jq1mod k Ž . next mini-slot, each B is moved on to the next A i.e., A in a cyclic j i iq1mod k Ž . manner see Fig. 10.5 . Since there are k output modules, a total of k mini-slots are necessary in each slot. Since the switch architecture uses output queuing, it has the best possible w x delay᎐throughput performance 5 . Consequently, we only need to compute the lost-cell probability. There are two sources of cell loss in the switch. First, cells are dropped if too many simultaneously arrive destined for an output Ž . group the loss due to knockout . Second, additional cells are dropped if the distributed, suboptimal routing algorithm cannot schedule a path through the w x switch. The analysis of this scheduling loss is complex and can be found in 5 . Here we just highlight the main results for the switch configuration without middle-stage trunking. Fig. 10.5 Parallel assignments in a cyclic manner from mini-slot to mini-slot. THE ATLANTA SWITCH