Design of bounded-degree circuits for parallel processing

The design of large circuits to interconnect many processors with low number of I/O ports and short diameter has been one of the major goals of researchers. Unfortunately, the underlying topologies of most of the popular circuits have a degree of the node which is a function of the network size. From the implementation viewpoint, networks with unbounded degree of the node pose two problems. First, there is a limit on the number of I/O channels allocated to a processor. Second, the I/O unit of the processor modules may need to be modified as the result of network expansion. In this paper, the design of high performance networks with constant degree of the node is addressed. A transformation applied to a circuit with a varying degree is essentially the replacement of each node with a topology with constant degree. The resulting hybrid circuit usually yields an efficient realization while preserving the advantageous features of designs with unbounded degree. New circuits such as: star-connected cycles, pancake-connected cycles, and bubble-sort connected cycles are introduced as examples of large circuits with constant degree of the node. The proposed networks lend themselves to an efficient VLSI implementation without compromising their efficiency in performing certain parallel algorithms.<
>