摘要： 高性能计算要解决的应用任务复杂多样，不同的应用任务具有不同的计算特征，在不同体系结构下计算效率差异巨大。为此，通过感知应用任务的计算-存储-通信(PMC)的资源需求，得到其PMC算粒，分配与其匹配的体系结构，使用超图描述应用任务程序结构和体系结构，利用超图同构原理构造超混合异构体系结构模型。实验结果表明，在应用任务的算粒感知驱动下，可重构体系结构模型具有计算效率高和能耗低的特点。
关键词:
高性能计算,
体系结构,
感知,
计算-存储-通信算粒,
超图,
算粒驱动
Abstract: Application tasks are complex and diverse, which need High Performance Computing(HPC) to solve. To solve the problem that computing efficiency is huge different when they run in different architecture because of different application tasks having different computing features. The Processor-Memory-Communication(PMC) resource requirements of application tasks are perceived to get PMC operator grain, the matched architecture is assigned to the applications. Hypergraph is used to describe the application structure and architecture, and hypergraph isomorphism principle is used to build super-mixed heterogeneous architecture. Experimental results show that reconfigurable architecture which operator grain perception driven has characteristic of high computing efficiency and low energy consumption.
Key words:
High Performance Computing(HPC),
architecture,
perception,
Processor-Memory-Communication(PMC) operator grain,
hypergraph,
operator grain driven
中图分类号: