模型指导的多维GPU软件低功耗优化方法

Model-Driven Multi-Dimensional Low-Power Optimization Method for GPU

作为众核体系结构的典型代表,GPU(Graphics Processing Units)芯片集成了大量并行处理核心,其功耗开销也在随之增大,逐渐成为计算机系统中功耗开销最大的组成部分之一,而软件低功耗优化技术是降低芯片功耗的有效方法.文中提出了一种模型指导的多维低功耗优化技术,通过结合动态电压/频率调节和动态核心关闭技术,在不影响性能的情况下降低GPU功耗.首先,针对GPU多线程执行模型的特点,建立了访存受限程序的功耗优化模型;然后,基于该模型,分别分析了动态电压/频率调节和动态核心关闭技术对程序执行时间和能量消耗的影响,进而将功耗优化问题归纳为一般整数规划问题;最后,通过对9个典型GPU程序的评测以及与已有方法的对比分析,验证了该文提出的低功耗优化技术可以在不影响性能的情况下有效降低芯片功耗.

As a typical many-core processor,GPU（Graphics Processing Units） integrates tremendous parallel processing cores,and the power consumption increases correspondingly,which makes it as one of the largest power consumers in modern computer systems.Software low-power optimization method is an effective method to reduce power consumption.This paper proposes a model-driven multi-dimensional low-power optimization method of coordinating dynamic voltage/frequency scaling and dynamic concurrency throttling method to lower the power consumption without sacrificing performance.Firstly,we establish the power optimization model targeted for memory-bounded program,based on the multi-thread execution model on GPU.Then,we analyze the impacts of dynamic voltage/frequency scaling and dynamic concurrency throttling on the program performance and energy consumption according to the established model,and induce the optimization problem as a typical integer programming problem.Finally,through detailed evaluation on 9 typical GPU applications and comparison with existing method,the experimental results validate the proposed multi-dimensional low-power optimization method could effectively reduce the energy consumption without performance loss.