The effect of grouped adaptive routing has been studied on the Clos network. Figure 10 shows the average network latency versus the network throughput for a 256 node 3-stage Clos network under random traffic with 32 byte packets. Deterministic routing and grouped adaptive routing are compared. Using grouped adaptive routing gives about 15% better maximum network throughput as well as slightly lower average latencies. This is because of the better utilisation of the links to the centre stage of the Clos network when using grouped adaptive routing.
The STC104 packet switch also supports universal routing, the routing algorithm on the 2-dimensional grid is as follows: as a packet enters the network it is first routed along the x-dimension to an intermediate destination switch which is chosen at random. The random header is deleted there and routing proceeds as in the deterministic case. Universal routing has the effect of distributing the traffic evenly over the entire network, thereby reducing worst case latencies.
Figure 11 shows the average latency of a 400 node 2-dimensional grid under random traffic as a function of the network throughput for universal and deterministic routing. It can be seen that the network saturates much earlier with universal routing, and that the latency increases much faster. This is because links some links have to be reserved for the random phase, so that in the x-dimension, the width of the link groups is only two and not four links. This causes greater contention and hence an increase in latency.
Figure 11: Universal routing on a 400 node 2-dimensional
Figure 10: Deterministic and grouped adaptive routing on a 256 node Clos network