用线程级推测技术在多核体系结构上并行化科学计算应用

Exposing Thread-level Speculation Parallelism in Scientific Applications on Multicore Architecture

线程级推测技术使在多核上加速传统上难以手工或自动并行化的串行程序成为可能,它不仅需要合理地选择线程的划分策略,而且需要合理地选择适合推测执行的应用.已有的大量研究主要集中在如SPEC CPU这样的桌面应用领域,为了全面地认识TLS技术的应用适用性,本文探讨TLS技术对科学计算应用的性能提升潜力,提出一套TLS适用性的基本判定准则,实验结果表明采用该技术加速SPLASH2中的多数应用可以有效利用16核及以上的计算资源.

Thread-level speculation （TLS） provides an opportunity for speeding up the serial codes on the multicore chips which are traditionally hard to parallel manually or automatically. It requires not only reasonable thread partition strategy,but also the suitable applications. Up to now,lots of researches have been mainly concentrated in the desktop applications,such as the SPEC CPU benchmark suite. In order to fully understand the TLS applicability,this paper explores the TLS potential parallelism in the scientific field,analyzes its impacting factors,and proposes a criterion for selecting the region to be speculatively executed. The experimental results show that the majority of SPLASH2 applications can make an effective use of more than 16 cores＇ computing resources in the TLS technology.