Evaluation of various node configurations for fine-grainmultithreading on stock processors

摘要

It becomes more and more interesting to construct multi-threadedparallel machines using stock processors due to their highperformance/price ratio. However no quantitative analysis has beenreported on the effectiveness of various node configurations and itsimpact on the overall performance. In this paper we explore threedifferent node configurations in detail and compare their dynamiccharacteristics through instruction-level simulation with six benchmarkprograms. Our experiments show that employing a dedicated processor forcommunication and synchronization is a reasonable approach because itcan almost double the performance. Several factors that limit theoverall speedup are also presented