A simulation-based comparison of interconnection networks

The authors evaluate some of the interconnection networks and routing algorithm proposed for large scale parallel computers. They compare different algorithms on networks built using a limited number of packages, each with a limited pin count, and using limited board layout area. Performance of each network is measured by simulating the execution of a synthetic program. The results quantify the extent to which network latency can be hidden because of the ability to initiate several memory accesses without blocking. The results also indicate that message combining is useful to avoid performance degradation in the presence of hot-spot traffic