论文部分内容阅读
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.
The problem of power (frequency) allocation is one of the most popular hotspots in the core processors.According to the performance-power model, we know its distribution in the auditory-frequency space, And the frequency level of these two dimensions through the implementation of step by step search, you can get “audit - frequency ” configuration of the optimal solution, so as to achieve the purpose of energy optimization; However, the existing methods in the field of audit - frequency space In order to solve this problem, a classical mathematical method based on solving the optimization problem, ie, the lowest energy consumption method of feasible direction method (energy- efficient optimization based on feasible direction method (EOFDM). At the same time, the search space can be simultaneously reduced from two dimensions of auditory and frequency, and converge to the lowest energy consumption point in iterative execution. The optimal heuristic hill-climbing heuristic (HCH) experiments were compared, the average execution times, execution time and energy consumption decreased by 39.5%, 46.8% and 48.3% respectively, which improved the convergence rate and reduced When the number of audits doubled, the average number of executions, execution time and energy consumption decreased by 48.8%, 51.6% and 50.9% respectively; when the frequency series doubled, the average execution times, execution time and energy consumption were respectively A decrease of 45.5%, 49.8% and 54.4%, respectively, resulting in improvements in convergence speed, search overhead and scalability.