一种基于推测代价评估的推测多线程并行粒度调节方法

A PARALLEL GRANULARITY TUNING APPROACH FOR SPECULATIVE MULTITHREADING BASED ON SPECULATIVE COST EVALUATION

传统的推测多线程技术总是假定程序的并行粒度大小应该随着处理器核资源数目的增加而增大,未考虑不同数目的处理器核资源对程序自身并行性能的影响作用。针对这个问题,提出一种自适应的循环并行粒度调节方法用于优化处理器核资源的分配过程。以推测级为单位,通过动态收集循环中所有推测线程的性能量化分析结果,进行推测代价评估。并利用评估结果动态调整循环的并行粒度大小,优化所分配到的处理器核资源的数目,以减少不必要的推测代价。实验表明,该方法不但在SPEC CPU基准测试程序集上能取得较好的性能提升,而且进一步优化了推测时的能耗开销。

Traditional speculative multithreading always assumes that the size of program's parallel granularity should increase as the number of processor core resources increases. It doesn't consider the effect of different number of processor core resources on the parallel performance of a program. Therefore, we proposed a self-adaptive parallel granularity adjustment for loops to optimize the allocation of their processor core resources. This approach took the speculative level as the unit, and performed the speculative cost evaluation by mean of dynamically collecting the results of performance quantitative analysis for all speculative threads within a loop. The results of cost evaluation were used to dynamically adjust the size of loop's parallel granularity and optimize the number of their allocated processor core resources to reduce the unnecessary cost for speculation. The experimental results show that our approach not only achieves better performance on SPEC CPU benchmark assemblies, but also optimizes the power consumption for speculation.