一种实时细颗粒度片上网络功率分配方法  

Real-time fine-gained power budget method for network-on-chip

作　　者：饶东宁[1] 李建华[1] 

机构地区：[1]广东工业大学计算机学院,广州510006

出　　处：《计算机应用研究》2016年第8期2358-2362,共5页Application Research of Computers

基　　金：国家自然科学基金资助项目(61100134);中央高校基本科研业务费专项资金资助项目(21615438)

摘　　要：片上网络(No C)不仅直接影响众核芯片的性能,而且消耗了芯片部分功率,在给定功率条件下使得片上网络的性能最优已经成为迫切需求。片上网络中路由器的工作负载(包的到达率)差异很大,需要对每一个路由器进行异构调整功率才能得到很好的性能。为此,针对如何在给定功率条件下优化No C性能进行了研究,提出了一个实时细颗粒度的功率分配方法,它能够根据每一个路由器的工作负载实时快速地分配功率,实现片上网络性能最优。实验结果表明相比其他功率分配方法,该方法平均减少26.1%的应用执行时间且具有较低的硬件开销。Network-on-chip （NoC） not only directly influences the overall many-core performance, but also consumes a sig- nificant portion of the total chip power. How to obtain the most optimal performance under the conditions of limited power has become an urgent need. The routers＇ workload （the packet arrival rate） in NoC is vary difference, so needs to adjust the power for each router heterogeneously to get good performance. Thus, this paper proposed a method of real-time fine-gained power budget in NoC under the limits of power which can perceive the touters＇ workload to distribute the power online and quickly to obtain the most optimize performance. Experimental results show that compared several state-of-the-art methods, the proposed method can reduce 26.1% of the application execution time averagely.

关 键 词：众核芯片 片上网络 功率受限 性能最优 功率分配 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]