Effective Prefetching and Replacement Policies in the Scalable Clustering-Based Multiprocessor System Design

Abstract

Recently, advancement in VLSI and packaging technologies demonstrates attractiveness in building scalable parallel systems using clustering-based architecture by exploiting communication locality. In this paper, we propose effective prefetching and replacement policies for our clustering-based multiprocessor system to enhance its total performance substantially. Our inter-clustering prefetching mechanism not only prefetches data for the single processor locality but also for the communication locality of processors in the same cluster by using inter-clustering caches. It also enhances conventional prefetching schemes by prefetching more data with less traffic overhead. Replacement policy proposed decreases replacement overhead effectively in the intra-clustering bus-based system. Simulation results show that prefetching and replacement schemes presented improve clustering-based system performance up to 33% and 25% respectively for the SPLASH[1] benchmark suites.