ParaInsight: An Assistant for Quantitatively Analyzing Multi-granularity Parallel Region

With the growing number of cores on chip, scalability of application becomes one of the most important metrics to impact performance and efficient utilization of hardware resource. Simple multi-threading transformation usually acts not well enough so that many of multi-threaded programs should be reconsidered for exploiting more parallelism. However, complex data dependency makes programmer hard to discover the regions with good scalability. Meanwhile, a variety of parallel programming models and platform properties make the choice of parallel granularity even difficult. In this paper, we develop a tool, named Para Insight, to help programmer obtain a quantification insight of parallelism on multi-granularity, as an assistant to ease the process of parallel analysis. We concentrate on revealing the inherent characteristic of program itself, independent of programming model and platform. While regions on which parallelization efforts should be focused are ordering, programmer could choose interested candidate and then use the quantification results to estimate the parallel zing overhead according to concrete parallel programming model and platform parameters. In contrast to many other parallelism discovery tools which mainly focus on parallel zing legacy serial programs, our tool provides assistance on the process of multi-threaded program development. We use multi-threaded benchmark suite Parsec and Splash2 to demonstrate the use of our tool, and specially use two cases to present how to improve performance of multithreaded application with the utilization of analysis data.