考虑负载差异性的多核处理器稳态温度及性能分析方法研究

发布时间：2018-05-12 09:46

本文选题：多核处理器 + 稳态分析　；参考：《哈尔滨工业大学》2017年博士论文

【摘要】：目前,温度及温度约束条件下的性能分析已经成为多核处理器设计自动化的重要组成部分。对于采用动态温度管理技术的多核处理器,处理器温度以及温度约束条件下的处理器性能高度依赖于处理器运行电压、频率和活动核个数。为了优化多核处理器的设计方案,需要分析处理器运行电压、频率、活动核个数与处理器温度及性能的关系。当处理器执行不同应用任务时,以及在执行同一个应用任务的不同阶段,处理器的负载具有很大的差异性。但到目前为止,现有以优化动态管理方案为目标的稳态分析方法都忽略了处理器负载的差异性。根据现有稳态模型,不能分析负载差异性对稳态温度及性能的影响,这必然会严重影响稳态分析的准确性。本文深入研究了多核处理器稳态温度以及温度约束条件下多核处理器性能的分析方法,用于分析处理器运行电压、频率、活动核个数与稳态温度及性能的关系。为了提高稳态分析的准确性、分析负载差异性对处理器温度及性能的影响,在温度及性能的分析模型中引入负载特征,来表示负载差异性。本文的主要研究内容包括:(1)提出了一种考虑负载差异性的多核处理器稳态温度模型为了提高多核处理器稳态温度分析的准确性,采用单周期各类型指令个数作为负载特征,来表示负载差异性,提出一种考虑负载差异性的多核处理器稳态温度模型。首先,将处理器核的微结构级动态功耗建模为单周期各类型指令个数、频率和电压平方的线性函数,将处理器核的微结构级静态功耗近似为电压与温度的线性函数。然后,通过理论推导,将处理器核的微结构级稳态温度表示为单周期各类型指令个数、频率、电压和活动核个数的函数。最后,通过仿真实验,验证了稳态温度模型的有效性,详细分析了频率、活动核个数与稳态温度之间的关系以及调整频率的核个数对稳态温度的影响。(2)提出了一种针对处理器热点选择的统计分析方法处理器最热的功能单元为热点。为了分析负载特征与处理器热点选择之间的关系,提出了一种针对处理器热点选择的统计分析方法。首先,基于考虑负载差异性的多核处理器稳态温度模型,提出了一种针对处理器热点选择的统计分析模型。给定一个负载特征子空间,使用该模型可以计算出某个功能单元为热点的概率。然后,基于提出的统计分析模型,给出了热点选择的统计分析算法,用于分析每个功能单元为热点的概率以及不同类型指令对热点选择概率的影响。最后,通过仿真实验,分析了处理器热点选择与负载特征之间的关系。(3)提出了一种性能约束多核处理器最小峰值温度的分析方法为了减少多核处理器最小峰值温度的分析时间,并且分析负载特征与处理器最小峰值温度的关系,针对基于帧的实时任务模型,提出一种基于解析法的多核处理器最小峰值温度分析方法。首先,将实时任务的单周期指令个数和各类型指令比例引入到多核处理器稳态温度的分析模型中。然后,针对基于帧的实时任务模型,提出多核处理器的最小峰值温度模型,将最小峰值温度表示为处理器运行电压、任务周期、任务指令个数、各类型指令比例和活动核个数的函数。最后,通过仿真实验,验证了所提分析方法的有效性,分析了负载特征、活动核个数与最小峰值温度之间的关系。(4)提出了一种温度约束多核处理器最大吞吐量的分析方法为了提高多核处理器性能分析的准确性、分析负载差异性对多核处理器性能的影响,在温度约束条件下,提出了一种考虑负载差异性的多核处理器最大吞吐量的分析方法。首先,通过理论推导,将多核处理器稳态温度表示为各类型指令比例、吞吐量、电压和活动核个数的函数。然后,证明了多核处理器吞吐量达到最大值的必要条件,并建立了多核处理器最大吞吐量的分析模型。最后,通过仿真实验,验证了所提分析方法的有效性,分析了各类型指令比例、单周期指令个数、活动核个数与多核处理器最大吞吐量之间的关系。(5)提出了一种多核处理器稳态温度及运行频率的概率分析方法为了对处理器的温度及运行频率进行概率分析,采用单周期指令个数作为负载特征,在单周期指令个数服从正态概率分布的条件下,提出一种多核处理器稳态温度及运行频率的概率分析方法。首先,在微结构级,将处理器的动态功耗建模为单周期指令个数和运行频率的线性函数,将静态功耗近似为温度的线性函数,将错误因子引入动态功耗和静态功耗的线性模型。然后,通过理论推导,将活动核和休眠核的热点温度表示为所有活动核单周期指令个数的线性函数。基于所有活动核单周期指令个数都服从相同正态概率分布的假设,通过理论分析,证明活动核和休眠核的热点温度服从正态分布。针对零减速调度策略,通过理论分析,得出处理器运行频率的概率分布。最后,通过仿真实验,分析了多核处理器稳态温度及运行频率的概率分布。
[Abstract]:At present, performance analysis under the constraints of temperature and temperature has become an important part of the design automation of multi-core processors. For multi-core processors with dynamic temperature management technology, processor performance is highly dependent on the operating voltage, frequency and number of active cores at the processor's temperature and temperature constraints. To optimize the design of multi core processors, it is necessary to analyze the relationship between the processor's operating voltage, frequency, the number of active cores and the temperature and performance of the processor. When the processor performs the different application tasks, and at different stages of the execution of the same application task, the processor's load is very different. The steady state analysis method of the dynamic management scheme ignores the difference of the processor load. According to the existing steady state model, it is impossible to analyze the influence of the load difference on the steady state temperature and performance. This will inevitably affect the accuracy of the steady state analysis. This paper has studied the steady-state temperature and temperature constraints of the multi core processor. The analysis method of multi-core processor performance is used to analyze the relationship between the operating voltage, frequency, the number of active cores and the steady state temperature and performance. In order to improve the accuracy of the steady state analysis, the influence of load difference on the temperature and performance of the processor is analyzed. The load characteristics are introduced in the analysis model of temperature and performance to represent the load difference. The main contents of this paper are as follows: (1) a steady-state temperature model of multi core processor, which considers load difference, is proposed to improve the accuracy of the steady-state temperature analysis of multi-core processors. The load difference is expressed by the number of single cycle types of instructions as the load characteristics, and a multi core processing considering the difference of load is proposed. First, the dynamic power consumption of the processor core is modeled as the linear function of the single cycle type instruction number, frequency and voltage square, and the micro structure level static power of the processor core is approximated to the linear function of the voltage and temperature. Then, the microstructural level steady temperature of the processor core is deduced by theory. The function of the number, frequency, voltage and number of active nuclei of each type of single cycle is expressed. Finally, the validity of the steady-state temperature model is verified by simulation experiments. The relationship between the frequency, the number of active nuclei and the steady temperature and the influence of the number of the adjusted frequencies on the steady temperature are analyzed in detail. (2) a kind of treatment is proposed. In order to analyze the relationship between the characteristics of the load and the hot spot selection of the processor, a statistical analysis method for the selection of the processor hot spots is proposed to analyze the relationship between the load characteristics and the hot spot selection of the processor. A statistical analysis model of the selection of the hot spot is given. Given a load feature subspace, the model can be used to calculate the probability of a functional unit as a hot spot. Then, based on the proposed statistical analysis model, a statistical analysis algorithm for hot spot selection is given, which is used to analyze the probability of each power unit as a hot spot and the different types of instructions. The influence on the probability of hot spot selection. Finally, through the simulation experiment, the relationship between the processor hot selection and the load characteristics is analyzed. (3) a performance constrained multi-core processor minimum peak temperature analysis method is proposed to reduce the minimum peak temperature analysis time of the multi-core processor, and analyze the load characteristics and the minimum processor. The relationship of peak temperature, a method of analyzing the minimum peak temperature of multi core processor based on analytic method is proposed for real-time task model based on frame. First, the number of single cycle instructions and the proportion of various types of instructions are introduced into the steady-state temperature analysis model of the multi core processor. Then, the real-time task based on frame is aimed at the frame based real-time task. The minimum peak temperature model of multi core processor is proposed, and the minimum peak temperature is expressed as the function of the processor operating voltage, task cycle, number of task instructions, the proportion of various types of instructions and the function of the number of active nuclei. Finally, the effectiveness of the proposed analytical method is verified by simulation experiments. The relationship between small peak temperature. (4) a temperature constrained multi-core processor maximum throughput analysis method is proposed to improve the accuracy of multi-core processor performance analysis and analyze the effect of load difference on the performance of multi core processors. Under the condition of temperature constraints, a multi core processor that considers load difference is proposed. First, through theoretical deduction, the steady-state temperature of the multi core processor is expressed as a function of the proportion of various types of instructions, throughput, voltage and the number of active nuclei. Then, the necessary conditions for the maximum throughput of the multi-core processor are proved, and the analysis model of the maximum throughput of the multi core processor is established. Finally, through the analysis of the maximum throughput of the multi core processor, the analysis model of the maximum throughput of the multi core processor is established. The simulation experiments verify the effectiveness of the proposed method, and analyze the relation between the proportion of instructions, the number of single cycle instructions, the number of active core and the maximum throughput of the multi core processor. (5) a probability analysis method for the steady-state temperature and running frequency of the multi-core processor is proposed in order to carry out the temperature and operation frequency of the processor. In probability analysis, a single period instruction number is used as a load feature. A probability analysis method for the steady-state temperature and operating frequency of a multi-core processor is proposed under the condition that the number of single periodic instructions obey normal probability distribution. First, the dynamic power consumption of the processor is modeled as a single period instruction number and running frequency at the micro structure level. The static power is approximated as the linear function of the temperature, and the error factor is introduced into the linear model of dynamic and static power consumption. Then, the hot temperature of the active and dormant nuclei is expressed as a linear function of the number of the single periodic instructions of the active kernel. The hypothesis of the same normal probability distribution, through theoretical analysis, proves that the hot temperature of the active core and the dormant kernel obeys the normal distribution. According to the zero deceleration scheduling strategy, the probability distribution of the operating frequency of the processor is obtained through theoretical analysis. Finally, the probability distribution of the steady-state temperature and operating frequency of the multi-core processor is analyzed by the simulation experiment.

【学位授予单位】：哈尔滨工业大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP332

【相似文献】