一种基于并行度分析模型的GPU功耗优化技术

A GPU Low-Power Optimization Based on Parallelism Analysis Model

随着硬件功能的不断丰富和软件开发环境的逐渐成熟,GPU开始被应用于通用计算领域,协助CPU加速程序的运行.为了追求高性能,GPU往往包含成百上千个核心运算单元.高密度的计算资源,使得其在性能远高于CPU的同时功耗也高于CPU.功耗问题已经成为制约GPU发展的重要问题之一.DVFS技术被广泛应用于处理器的低功耗优化,而对GPU进行相应研究的前提是对其程序运行过程进行分析和建模,从而可以根据应用程序的特征来确定优化策略.此外,GPU主要由图形处理器芯片和片外的DRAM组成,有研究指出针对这类系统的功耗优化应当综合考虑处理器和存储器,使二者可以互相协调以达到更好的优化效果.文中在一个已有的基于程序并行度分析的GPU性能模型的基础上,综合考虑计算部件与存储部件的功耗,建立了性能约束条件下的GPU功耗优化模型.对于给定的程序,在满足性能约束的前提下,以功耗最优为目标分别给出处理器和存储器的DVFS优化策略.作者选取了9个测试用例在3种模拟平台上进行了实验验证,结果表明文中的方法可以在满足性能约束条件10%的误差范围内获得最优的GPU能量消耗.

With the continues development of hardware and software,GPU has been used in general purpose computation field,accelerating applications for CPU.To achieve high computing performance,GPU typically includes hundreds of computing units.The high density of computing resource on chip brings in high power consumption as well as high performance.The power consumption problem has become one of the most important problems for the development of GPU.The DVFS technique is widely used to optimize power consumption for processors.However,applying the DVFS technique to GPU depends on the analysis of program execution on GPU,so that optimization strategy can be chosen according to the program feature.Besides,GPU is comprised of a processor chip and an off-chip DRAM system.Some previous researches point out that the power consumption optimization for such a system should involve both the processor and the DRAM,to achieve better optimization effect.Based on an existing GPU analytical model,this paper proposes a GPU power optimization model under performance restriction,involving both the processor and the DRAM on GPU.For a given program,the model gives the DVFS strategies for the processor and the DRAM respectively with an appointed performance restriction.The authors choose nine test cases to evaluate the model on three simulated GPU platforms.The experimental results show that the model can achieve optimal energy consumption while the performance deviation from the restriction is less than 10%.