The scalability of linear filters on hypercube concurrent computers

The general concept in solving a problem on a parallel machine requires subdividing a large problem into processes, each of which can be computed on a single processor. This can be a formidable task requiring substantial communication and control overhead. It is easy to speculate what may happen when one implements various applications on such a system. But nothing teaches one as surely as actually implementing real problems with real software on real hardware. This paper reports on the implementation and performance of a hypercube arrangement of VLSI DSP processors (TMS320C30) for FIR filtering. Two approaches are considered. A direct partitioning of the standard FIR filter structure and a decomposition of the fast FIR algorithm described by Mou and Duhamel (1987). Two interprocessor communication strategies are described. Models constructed from measurements on a prototype machine are used to calculate the performance of filters on higher dimensional hypercubes