LODS: Locality-oriented dynamic scheduling for on-chip multiprocessors

Current multiprocessor SoC applications like network protocol codes, multimedia processing and base-band telecom circuits have tight time-to-market and performance constraints, which require an efficient design cycle. Consequently, automated techniques such as those oriented towards exploiting data locality are critical. In this paper, we demonstrate that, existing loop scheduling techniques provide performance improvements even on on-chip multiprocessors, but they fall short of generating the best results since they do not take data locality into account as an explicit optimization parameter. Based on this observation, we propose a data locality-oriented loop-scheduling algorithm. The idea is to assign loop iterations to processors in such a fashion that each processor makes maximum reuse of the data it accesses.