摘要
应用映射是片上网络体系结构研究的关键问题之一,映射结果的好坏会极大地影响体系结构的性能。现有的应用映射方法大多基于特定的网络结构,如2d-mesh、2d-torus等,研究NoC性能或功耗约束的应用映射与优化方法。本文提出了一种拓扑结构感知的基于高层代码转换的片上网络应用映射与优化方法。该方法采用多面体模型对应用的核心循环进行自动并行和局部性优化,并将网络拓扑结构抽象成带权重的有向图,使用该有向图对任务流图进行覆盖,以提高任务的并行性,降低任务间同步和通信开销。实验结果表明,采用优化的映射方法后任务节点间的并行性被充分利用,通信开销降低,整体上提高了片上网络系统性能。
Application mapping is a key problem in the research area of NoC architecture. The result of mapping has greatly impact on the performance of the architecture. Mapping methods proposed recently are mainly focused on bandwidth or energy constraints for regular NoC architectures, such as 2d-mesh, 2d-torus, etc. In this paper, we proposed a topologyaware application mapping method for NoC architecture based on high level application transformation. Polyhedral model is adopted to optimize the kernel loop nests with parallelization and data locality. We abstract the topology of NoC to a weighted directed topology graph, and used the graph to map and optimize application tasks. Experiment results show that with our method, the parallelization of the tasks is fully used and communication cost between tasks is reduced, so the performance of the NoC system is improved and energy consumption is reduced.
出处
《计算机工程与科学》
CSCD
北大核心
2009年第A01期109-111,167,共4页
Computer Engineering & Science
基金
国家自然科学基金资助项目(90707003)
关键词
片上网络
应用映射
拓扑结构
多面体模型
高层转换
network-on-chip
application mapping
topologys
polyhedral model
high level transformation
