一种面向非规则数据的阶段预取策略

A Phased Prefetching Scheme for Irregular Data References

传统数据预取技术在处理结构复杂的非规则数据应用程序时,其有效性明显下降.为解决该问题,基于程序运行时的数据访问阶段性特征,提出一种面向非规则数据的阶段预取策略,研究应用程序的访存规律和预取调度机制.该策略通过在线剖析应用程序的访存行为,识别出数据访问性能指标表现稳定的数据访问阶段和具有特定访存行为特征的预取阶段,实现在数据访问阶段内依据预取阶段的访存规律动态调整预取操作.实验结果表明,与传统的基于访存流模型的数据预取技术相比较,阶段预取策略能够减少无用预取,更加有效地改善非规则数据应用程序性能.

When dealing with the structurally complex irregular data applications, the effectiveness of conventional data prefetching techniques reduces. To solve this problem, a new phased prefetching technique for irregular data is presented based on memory access stream. Phased prefetching mechanism recognizes data access phases and prefetching phases by profiling programs at runtime, and adjusts prefetch operations accordingly. Experimental results show that phased prefetching mechanism can decrease useless prefetches and improve system performance of irregular data applications more effectively compared with conventional data prefetching techniques.