暗硅时代多核系统资源管理算法研究

发布时间：2018-02-27 01:19

本文关键词： 暗硅热设计功耗多核系统动态规划应用映射模拟退火反馈调整　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：集成电路制造工艺的发展减小了晶体管的特征尺寸,理论上单位面积的芯片可集成更多晶体管而功耗密度保持不变。但是,当工艺节点发展到22nm以下,泄漏功耗开始主导晶体管功耗,并随着工艺节点的缩小呈指数级增长,这引发了芯片的过热问题。业界预测:当工艺节点继续往前发展,芯片上将会有更多的功能模块不能同时处于全频率工作状态,即总是部分处于开启而其他处于关闭状态,这预示着暗硅时代的到来。单核系统的局限性催生了多核系统的发展,而集成电路向暗硅发展的趋势使系统中的处理器必须合理的开断和配置以缓解热问题。多核系统的资源管理算法可根据应用自身的特点,对芯片中热设计功耗,处理器的开启数目和工作频率等资源进行合理分配,满足温度、系统资源约束的同时优化系统性能。本文在暗硅背景下,面向于共享式存储结构的同构多核系统,提出了一种资源管理算法。首先,针对具有线程并行性的应用集,根据其在不同处理器数目和工作频率的功耗及吞吐特性,利用动态规划配置处理器数目和工作频率,实现当前热设计功耗约束下的系统吞吐最优化;其次,以提高散热效果和降低存取代价为目标,使用模拟退火算法完成应用映射,确定处理器开断及应用在系统中的布局,减小过热点出现的几率,并根据应用布局,通过温度仿真获得系统温度分布;最后,根据有无过热点的反馈,充分利用系统温度裕度,循环迭代地调整热设计功耗大小,每次调整后重新进行资源配置和应用映射,最终在最大热设计功耗下获得系统最优性能。本文搭建了系统吞吐、功耗及温度的仿真环境,使所提资源管理算法可嵌入于该环境中一体化完成。文中借助多核架构仿真工具模拟应用在同构系统中的执行过程,可获得应用在不同处理器数目及工作频率下的吞吐特性。提取系统架构参数和运行信息,并转换为功耗仿真工具的输入文件,便可获得应用功耗特性。以吞吐和功耗特性为输入,算法在调整热设计功耗的过程中,使用温度仿真工具进行热仿真,并能获得温度布局。所搭建环境可灵活地应用于同构多核系统,实验表明,所提调度方法能够有效的避免过热点,并优化了系统性能,且相比于棋盘式布局,系统最高温度降低3%,相比开关式调整过热点的方法,系统吞吐量最大增加约12%。
[Abstract]:The development of integrated circuit manufacturing process reduces the characteristic size of transistors. In theory, more transistors can be integrated into chips per unit area and the power density remains constant. Leakage power starts to dominate transistor power consumption, and increases exponentially as the process node shrinks, causing the chip to overheat. There will be more modules on the chip that can't work at full frequency at the same time, that is, always partially on and off, which bodes well for the advent of the dark silicon era. The limitations of mononuclear systems have spawned the development of multicore systems. The trend of the development of integrated circuits towards dark silicon makes the processors in the system must be switched on and configured reasonably to alleviate the heat problem. The resource management algorithm of multi-core systems can consume the power of thermal design in chips according to the characteristics of the applications. The open number and working frequency of processors are allocated reasonably to satisfy the temperature and system resource constraints and to optimize the system performance. In this paper, an isomorphic multi-core system with shared memory structure is proposed in the background of dark silicon. In this paper, a resource management algorithm is proposed. Firstly, for the application set with thread parallelism, according to the power consumption and throughput characteristics of different processor numbers and working frequencies, dynamic programming is used to configure the number and working frequency of processors. In order to improve the heat dissipation effect and reduce the access cost, simulated annealing algorithm is used to complete the application mapping to determine the switch on and the layout of the processor in the system. According to the application layout, the temperature distribution of the system is obtained by temperature simulation. Finally, according to the feedback of whether there is a hot spot, the system temperature margin is fully utilized, and the power consumption of thermal design is adjusted iteratively and iteratively. Resource configuration and application mapping are redone after each adjustment, and the optimal performance of the system is finally obtained under the maximum thermal design power consumption. In this paper, a simulation environment of system throughput, power consumption and temperature is built. So that the proposed resource management algorithm can be embedded in the environment to be completed. In this paper, the execution process of the application in the isomorphic system is simulated by means of the multi-core architecture simulation tool. It can obtain the throughput characteristics applied in different processor number and working frequency, extract system architecture parameters and operation information, and convert to input file of power emulation tool. With the input of throughput and power consumption characteristics, the algorithm uses the temperature simulation tool to conduct thermal simulation in the process of adjusting the power consumption of thermal design. The environment can be applied to isomorphic multi-core system flexibly. Experiments show that the proposed scheduling method can effectively avoid hot spots and optimize the performance of the system, and compared with the chessboard layout, the proposed scheduling method can be applied to the isomorphic multi-core system flexibly, and compared with the chessboard layout, the proposed scheduling method can effectively avoid the over-hot spots. The maximum temperature of the system is reduced by 3 and the maximum throughput of the system is increased by about 12 percent compared with the switching method of adjusting the hot spot.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TN405

【参考文献】