面向多GPU架构电流预测模型研究  

Research on Current Prediction Model for Different Architectures′GPUs

作　　者：赵德玉[1] 陈庆奎[1,2] ZHAO De-yu;CHEN Qing-kui(School of Management,University of Shanghai for Science and Technology,Shanghai 200093,China;School of Optical-Electrical and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)

机构地区：[1]上海理工大学管理学院,上海200093 [2]上海理工大学光电信息与计算机工程学院,上海200093

出　　处：《小型微型计算机系统》2022年第10期2051-2056,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61572325,60970012)资助;山东省重点研发项目(2018GGX101005)资助.

摘　　要：为降低GPU通用计算能耗分析的复杂性,提高GPU电流预测模型的通用性,提出了一种基于Elman神经网络的面向多GPU架构GPU运行电流通用预测模型.通过分析GPU程序,提取计算操作数量、存储访问操作数量和程序分支数量三个程序特征;引入GPU体系结构复杂度系数;最后,采用Elman神经网络构建了程序特征、体系结构复杂度与运行电流的关系模型.实验结果表明,单程序的预测误差不超过8%,不同体系结构间的平均预测误差不超过7%.该电流预测模型为深入分析GPU通用计算能耗复杂性和多程序并行调度奠定了良好的理论基础.In order to reduce the complexity of energy consumption analysis and improve the versatility of GPU current prediction model,a general current prediction model for different architectures′GPUs based on Elman neural network was proposed.By analyzing the GPU application,three program features including the number of calculation instructions,the number of storage instructions and the number of branch instructions were extracted.A GPU architecture complexity parameterwas introduced.The relationship model of program features,architecture feature and GPU running current was constructed by using Elman neural network.The experimental results showed that the prediction error of single program was less than 8%,and the average prediction errors of different architectures were less than 7%.The model laid a good theory foundation for GPU energy complexity analysis and multi-application parallel scheduling.

关 键 词：GPU 能耗分析 电流预测 ELMAN神经网络 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]