Characterization of Locality in Loop-Parallel Programs

Data locality is one of the most important characteristics of programs. Its study has significant influence on the future development of architectures and compilers. Shared-memory multiprocessor (SMP) machines and their applications have become widely available, but there are few studies in the classification of data locality in parallel programs. Most studies have focused on temporal and spatial locality, and false sharing as metrics by which to optimize cache-coherence actions. In this paper, we propose a classification of data locality in loop-based parallel programs on SMP machines. The classification is expressed in terms of the cacheline sharing, processor sharing, memory reference instruction reuse, and parallel/serial region sharing. This classification helps us understand the characteristics of complex parallel programs. For analysis, we have implemented a tool, called ParaSim which consists of two parts: a code augmenter and a memory simulator. It can augment highly optimized..