A Thread Partitioning Algorithm for Data Locality Improvement

摘要

This paper addresses the problem of automatic thread partitioning of sequential programs. We propose a new algorithm, which has the main features as follows: 1) non-loop regions of programs may be parallelized; 2) partitioning is performed in the way that data locality in the new threads is improved. The algorithm is preliminarily evaluated using one handcrafted example and a generator of data dependence graphs, tuned to generate random graphs with characteristics close to those of real-world programs, and a cache simulator. The results of evaluation are presented which shows feasibility of our approach.