Interconnects for high bandwidth, highly parallel distributed memory supercomputers

Presents a new interconnect topology, called TICNET, that combines the favorable blocking behavior of the hypercube with the economy of the 2D-mesh. The behavior of 2D-mesh, hypercube, and TICNET is evaluated by a statistical analysis by which connectivity and blocking behavior are investigated. The analysis answers important questions such as: how many logical connections can exist simultaneously, when will the network saturate, how well are the physical links utilized, and what is the cost of realization of the network. It is shown that the TICNET, which can be realized as a hierarchy of crossbars, is similar in behavior to the hypercube but much more cost-effective. Compared to the 2D-mesh, the TICNET has a much better blocking behavior and is still more cost-effective. Consequently, the authors see the hierarchy of crossbars (TICNET) as the preferred interconnect for future high bandwidth, highly parallel distributed memory architectures.<<ETX>>