New program model for program partitioning on NUMA multiprocessor systems

Abstract

A new program model is presented to accurately represent parallel programs for partitioning and scheduling problems. This model extends the graphic representation of the macrodataflow by considering the complex communication options supported by NUMA systems. The proposed model shows not only task precedence relations but also data sharing status. Moreover, a new partitioning method based on the proposed model is also developed. Experimental results show that many conventional partitioning algorithms operate more efficiently using the proposed model, and that the proposed algorithm surpasses existing algorithms.