Appendix V: Performance Metrics

Here are some results of how the simulated topologies performed under random traffic. Clearly the deBruijn(11) cycle is bottlenecked, but simply bisecting the cycle into a MDSXN[5,6] leads to acceptable performance. Of course the threshold at which a topology fails is dependent upon the amount traffic and the bandwidth of each link in the topology. Should the former increase or the latter decrease then the MDSXN[5,6] and MDSXN[3,4,4] will also begin to fail, so in cases of simulation of actual NUMA systems, these factors should be set as accurately as possible.
The SCC(6) topology also fails under this simulated traffic, and the SCI(6) illustrates that the bottleneck occurs in the network links and not in the lateral links. For all these tests I intended to route an average of 1K words of traffic per node in the topology, thus 2.048M words for the 2n node topologies and 3.600M for the (n-1)n! node topologies, but the SC topologies were bottlenecked with runs of only half this amount and threatened not to finish simulation of routing 3.600M words of data traffic for several days.