EOFDM:一种面向众核架构的最低能耗搜索方法

发布时间：2018-03-21 16:21

本文选题：能耗优化　切入点：众核　出处：《计算机研究与发展》2015年06期 　论文类型：期刊论文

【摘要】：面向能耗优化的面积(核数)-功率(频率)分配问题是当前众核处理器研究热点之一.通过性能-功耗模型了解其在核数-频率空间的分布规律,然后在核数和频率级别这2个维度上通过实测执行逐步搜索,可以获取"核数-频率"配置的最优解,从而达到能耗优化的目的;然而本领域现有方法在核数-频率空间内实测搜索最低能耗时收敛速度慢、搜索开销大、可扩展性差.针对此问题,提出了一种基于求解最优化问题的经典数学方法——可行方向法的最低能耗搜索方法(energy-efficient optimization based on feasible direction method,EOFDM),每次执行都能从核数和频率2个维度上同时减小搜索空间,在迭代执行中快速收敛至最低能耗点.该方法与现有研究中最优的启发式爬山法(hill-climbing heuristic,HCH)进行了对比实验,平均执行次数、执行时间和能耗分别降低39.5%,46.8%,48.3%,提高了收敛速度,降低了搜索开销;当核数增加一倍时,平均执行次数、执行时间和能耗分别降低48.8%,51.6%,50.9%;当频率级数增加一倍时,平均执行次数、执行时间和能耗分别降低45.5%,49.8%,54.4%,在收敛速度、搜索开销和可扩展性方面均有提高.
[Abstract]:The area (kernel-power (frequency)) allocation problem oriented to energy consumption optimization is one of the hot topics in the research of multi-core processors at present. The distribution law in the kernel-frequency space is understood by using the performance-power model. Then the optimal solution of the "Kernel Frequency" configuration can be obtained by performing step by step search on the two dimensions of the kernel number and the frequency level, so as to achieve the purpose of energy consumption optimization. However, the existing methods in this field have slow convergence speed, large search cost and poor scalability when the minimum energy consumption is measured in the kernel-frequency space. In this paper, a classical mathematical method for solving optimization problems, the feasible direction method, is proposed. The energy efficiency optimization based on feasible direction method EOFDM can reduce the search space from two dimensions of kernel number and frequency at the same time. The method is compared with the optimal heuristic heuristic hill-climbing heuristic (HCH). The average execution times, execution time and energy consumption are reduced by 39.5%, 46.8% and 48.3%, respectively, and the convergence rate is improved. The search cost is reduced; when the number of cores is doubled, the average execution times, execution times and energy consumption are reduced by 48.8and 51.6and 50.9, respectively; when the frequency series is doubled, the average execution times, execution times and energy consumption are reduced by 45.59.8and 54.4s, respectively, and at the convergence rate, Search cost and scalability are improved.
【作者单位】：计算机体系结构国家重点实验室(中国科学院计算技术研究所);中国科学院大学计算机与控制学院;河北农业大学信息科学与技术学院;国家计算机网络应急技术处理协调中心;北京交通大学计算机与信息技术学院;数学工程与先进计算国家重点实验室;
【基金】：国家“九七三”重点基础研究发展计划基金项目(2011CB302501) “核高基”国家科技重大专项基金项目(2013ZX0102-8001-001-001) 国家“八六三”高技术研究发展计划基金项目(2015AA011204,2012AA010901) 国家自然科学基金项目(61173007,61204047,61332009)
【分类号】：TP332

【参考文献】