Performance analysis of interconnection networks for massively parallel multicomputers

The authors present a performance analysis model and describe a network analysis program that can be used to estimate end-to-end delay and maximum system throughput for massively parallel multicomputer networks. The model explicitly accounts for both message processing and transmission delays. The model is general enough to allow analysis of message and virtual cut-through switching, point-to-point and multiple bus topologies, unidirectional and bidirectional links, and varying traffic locality assumptions. The model and program allow the analytical evaluation of network and node design alternatives, as illustrated by an application example