暗硅多核系统芯片资源调度算法

发布时间：2018-05-04 14:44

本文选题：暗硅 + 热设计功耗　；参考：《计算机辅助设计与图形学学报》2017年06期

【摘要】：芯片集成度的提升芯片带来功耗密度的增加,引起芯片的过热问题.近年来,人们提出暗硅设计的概念,有选择地关闭部分工作模块,避免芯片上所有模块同时处于开启状态,以解决过热问题.为此,提出一种基于模拟退火的多核系统资源调度算法.针对具体的应用采用迭代方法调整热设计功耗约束、分配处理器资源,并确定芯片模块的开启和关闭,在保证系统吞吐的同时,有效地解决芯片的过热问题.首先,针对已知应用集,在热设计功耗和系统约束下通过动态规划为每个应用配置处理器数目和频率等级.其次,基于模拟退火算法以散热效果和通信延迟为目标完成应用映射,确定开启和关闭的处理器.最后,根据有无过热点的反馈,迭代地调整热设计功耗大小,获得系统最大热设计功耗,并据此获得应用的最终资源配置和映射结果.所提调度方法能够有效地避免过热点,在资源约束下最优化系统性能.实验结果表明,相比于棋盘式布局,系统最高温度能够降低3%,相比开关调整过热点的方法,系统吞吐量能够最大增加约12%.
[Abstract]:The increase of chip integration leads to the increase of power density, which leads to the overheating of the chip. In recent years, the concept of dark silicon design has been proposed to selectively close some working modules so as to avoid all the modules on the chip being on at the same time, so as to solve the problem of overheating. A resource scheduling algorithm for multi-core systems based on simulated annealing is proposed. For specific applications, iterative method is used to adjust thermal design power constraints, allocate processor resources, and determine the opening and closing of chip modules, which can effectively solve the overheating problem of the chip while ensuring the system throughput. First, the number of processors and frequency levels for each application are dynamically programmed for a given set of applications under thermal design power consumption and system constraints. Secondly, based on simulated annealing algorithm, heat dissipation and communication delay are taken as the targets to complete application mapping to determine which processor to turn on and off. Finally, according to the feedback whether there is a hot spot, the power consumption of thermal design is adjusted iteratively to obtain the maximum thermal design power consumption of the system, and the final resource configuration and mapping results of the application are obtained. The proposed scheduling method can effectively avoid over-hot spots and optimize system performance under resource constraints. The experimental results show that compared with the chessboard layout, the maximum temperature of the system can be reduced by 3%, and the system throughput can be increased by about 12% compared with the method of switch adjusting the hot spot.
【作者单位】：中国科学技术大学电子科学与技术系;中国科学技术大学微纳电子系统集成研究中心;
【基金】：国家自然科学基金(61404123) 安徽省自然科学基金(1508085MF134)
【分类号】：TN40

【相似文献】