一种基于关键路径分析的CPU-GPU异构系统综合能耗优化方法

GPU强大的计算性能使得CPU-GPU异构体系结构成为高性能计算领域热点研究方向.虽然GPU的性能/功耗比较高,但在构建大规模计算系统时,功耗问题仍然是限制系统运行的关键因素之一.现在已有的针对GPU的功耗优化研究主要关注如何降低GPU本身的功耗,而没有将CPU和GPU作为一个整体进行综合考虑.文中深入分析了CUDA程序在CPU-GPU异构系统上的运行特点,归纳其中的任务依赖关系,给出了使用AOV网表示程序执行过程的方法,并在此基础上分析程序运行的关键路径,找出程序中可以进行能耗优化的部分,并求解相应的频率调节幅度,在保持程序性能不变的前提下最小化程序的整体能量消耗.

An Integrated Energy Optimization Approach for CPU-GPU Heterogeneous Systems Based on Critical Path Analysis

GPU’s powerful computing ability has attracted much attention in the high performance computing field currently,and the CPU-GPU heterogeneous architecture has become a hot research direction.Although the performance/watt ratio of the GPU is higher than general purpose CPU,the power consumption problem is still one of the critical problems in constructing high performance computing system with CPU and GPU.Most existed power optimization researches oriented to the GPU focus on decreasing the power consumption of the GPU solely.Few works have considered the energy optimization for the whole program involved the CPU and the GPU.In this paper,we analyze the execution characteristics of the CUDA program on the CPU-GPU heterogeneous system deeply,and induce the dependency relationships of tasks in the program in detail.Then,we describe how to express the execution of the program with AOV network,based on which we propose the energy optimization method.Our method analyzes the critical path of the program from the AOV network,detects the tasks that can be optimized for energy and calculates the frequency scaling factors to minimize the whole program’s energy consumption while maintaining the performance.