Parallelization Strategies for Mixed Regular-Irregular Applications on Multicore-Systems

Scientific simulation codes often exhibit a mixed structure of regular and irregular data accesses. Since the organization of data accesses has a large influence on the overall performance of parallel code, a careful planning of parallelism is required. In this article, we consider a mixed regular-irregular particle simulation code and investigate several parallelization strategies for multicore architectures consisting of several multicore processors in a shared memory system. The interaction of irregular and regular data accesses are the specific challenge for a cache optimized parallel multicore-code. We present performance experiments on three different multicore systems and show that a mixture of parallelization techniques for irregular and regular applications leads to the best performance.