A Comparative Performance Study of an Interconnection Cached Network

The Holy Grail of multi-computer network architects is an ideal crossbar with large size, high bandwidth and fast switching speed. Currently, only small crossbars with fast switching rates are available. Some technologies, e.g. optics, make it possible to build large, high bandwidth crossbars. However, these crossbars switch slowly. The Interconnection Cached Network (ICN) contains many small crossbars interconnected by a single large crossbar. The large crossbar is used for topology reconfiguration, and the small crossbars for message switching. In many parallel applications, each processing element (PE) communicates with only a small set of other PEs. We call this phenomenon switching locality and show that, in its presence, the ICN performs almost as well as the ideal crossbar. We also show that alternative networks have poorer overall performance in terms of average message latency, highest sustainable throughput, and application scalability.