Parallel task scheduling for the block data flow architecture

An important issue in the development of high performance computer architectures is how to automatically partition algorithms for parallel execution. The cyclostatic realization method (CRM) is identified as a basis for an effective mapping methodology to achieve automatic partitioning. The paper develops an advanced mapping methodology for the block data flow architecture (BDFA) based on CRM. Weaknesses in the existing methodology are identified and an effective, generalized methodology for correcting these weaknesses is proposed. Several key factors are investigated: a system-wide average iteration period for multiple input streams, the effect of communication overhead, and the influence of network topology. The enhanced mapping methodology is applied to the BDFA for a 2D 2nd order IIR filter.