基于数据依赖的GPU功耗管理方法研究  

Research on Power Management Method of GPU Based on Data Dependence

作　　者：魏雄 王秋娴 胡倩 闫坤 许萍萍 WEI Xiong;WANG Qiuxian;HU Qian;YAN Kun;XU Pingping(School of Mathematics and Computer,Wuhan Textile University,Wuhan 430000,China;School of Computer,National University of Defense Technology,Changsha 410073,China;Hubei Urban Construction Vocational and Technological College,Wuhan 430205,China)

机构地区：[1]武汉纺织大学数学与计算机学院,湖北武汉430000 [2]国防科技大学计算机学院,湖南长沙410073 [3]湖北城市建设职业技术学院,湖北武汉430205

出　　处：《计算机与网络》2021年第15期66-72,共7页Computer & Network

摘　　要：图形处理器(GPU)因其高并发和高吞吐量的特性被广泛应用于大数据和人工智能等高性能计算领域,随着超大规模集成电路技术的发展,片上集成的处理单元越来越多,高功耗在增加设备运行成本的同时,降低电池的使用时间和集成电路芯片的可靠性。针对功耗问题,提出一种基于数据依赖的GPU功耗管理方法(DDPM),通过优化线程分配和缓存置换策略减少GPU系统功耗。实验结果表明,DDPM相较于共享感知数据管理方法,L1缓存命中率提高了7%,DRAM数据传输量降低了8.1%;MC-aware-ORI,MC-aware-LoSe,MC-aware-SiOb方法能效分别提高了2.7%,2.65%,8.67%。Due to its high concurrency and high throughput,Graphics Processing Unit(GPU)is widely used in high-performance computing such as big data and artificial intelligence.With the development of VLSI technology,more and more processing units are integrated on chip.High power consumption increases the operating cost of equipment,while reducing the battery life and the reliability of integrated circuit chips.To solve the problem of power consumption,a data dependency based GPU power management method(DDPM)is proposed to reduce the power consumption of GPU system by optimizing thread allocation and cache replacement strategy.The experimental results show that compared with the shared sensing data management method,DDPM can improve the L1 cache hit rate by 7%,reduces DRAM data transmission capacity by 8.1%,and the energy efficiency of the MC-aware-ORI,MC-aware-LoSe,MC-aware-SiOb method is increased by 2.7%,2.65%and 8.67%respectively.

关 键 词：数据依赖 线程调度 缓存置换 功耗优化 

分 类 号：TP338.6[自动化与计算机技术—计算机系统结构]