面向异构多核并行系统的层次化计算模型HmPlogP

摘要

Based on the parameterized LogP (PLogP) model, this paper proposes a new hierarchical computation model named the HmPlogP for heterogeneous multi-core parallel systems. HmPlogP considers the characteristics of the systems such as the heterogeneity of conventional cores and accelerator cores, and the hierarchy of memory access and parallelization. The model abstracts the communication and the memory access of a heterogeneous multi-core parallel system in a hierarchical way, uses vectorization parameters to represent the characteristics of a parallel system in different levels, and makes prediction of the memory access overhead of accelerating cores to direct the design and the optimization of parallel programs. The experimental result shows that the HmPlogP model can effectively predict the communication time of conventional cores and the memory access time of accelerating cores.%在参数化LogP模型(PLogP模型)的基础上,针对异构多核并行系统通用核和加速核的异构性、存储的层次化、并行执行的层次化特征,提出了新的层次化计算模型Hm-PlogP.该模型对异构多核并行系统的通信和访存进行了抽象,采用向量化参数表达并行系统不同层次的特征,能够预测加速核的执行开销并以此指导并行程序的设计和优化.实验结果表明,HmPlogP模型能有效地预测通用核的点到点通信时间和加速核的访存时间.