Parallel Data Reuse Theory for OpenMP Applications

The development of multi-core processor technology makes parallel programming become more and more popular. Similar to serial programs on single-core platforms, the locality optimization of parallel programs is and will be a hot-spot of research owing to the memory wall problem. In this paper, we extend the famous data reuse theory to parallel domain and propose parallel data reuse theory for OpenMP applications. The parallel data reuse theory further classifies the reuse in parallel programs, from four classes to eight. This paper systemically discusses the intra-/inter-iteration reuse and intra-/inter-processor reuse in OpenMP programs, and gives the judging and solving method of each reuse class. Besides, this paper does the case study and analysis of SPEComp2001 benchmarks, using our parallel data reuse theory. We believe that parallel data reuse theory will have a big impact on the locality optimization of parallel applications.