Analysis Of Steady-state Temperature And Performance For Multicore Processors Considering Workload Variation

At present,the analysis of temperature and performance has become an important part of the design automation for multicore processors.For multicore processors using dynamic thermal management(DTM)techniques,the temperature and performance greatly depends the voltage,frequency and number of active cores.In order to optimize the design of multicore processors,it is necessary to analyze the relationship between voltage,frequency,number of active cores and temperature of processor.When a processor executes different tasks,and is at the different stages of the execution of a task,the workload of the processor is very different.But so far,the existing steady-state analysis methods aiming to optimize DTM schemes have neglected the workload variations.According to the existing steady-state model,it is unable to analyze the impact of workload variations on the steady-state temperature and performance,which will seriously affect the accuracy of steady-state analysis.In this dissertation,the method for analyzing the steady-state temperature and performance of multicore processors are deeply studied,which is used to analyze the relationship between voltage,frequency,number of active cores and steady-state temperature and performance.In order to improve the analyzing accuracy of the steady-state temperature and performance,and analyze the impact of workload variations on the steady-state temperature and performance of multicore processors,the workload characteristics representing the workload variations are introduced into the model of the steady-state analysis.The main contributions of this dissertation are as follows:(1)A steady-state thermal model of multicore processors considering workload variationsTo improve the analyzing accuracy of the steady-state temperature of multicore processors,the various instructions per cycle(VIPC)is adopted as the workload characteristic to represent the workload variations,and a steady-state thermal model of multicore processors considering workload variations is proposed.First,at the micro-architecture level,the dynamic power of a core is modeled as a linear function of the VIPC,running frequency and the square of voltage.The leakage power is approximated as a linear model of the temperature and voltage.Second,the micro-architecture-level steady-state temperature of a core is derived as a function of the VIPC,frequency,voltage and the number of active cores.Finally,the proposed steady-state thermal model is verified by the experiment.The relationships between the frequency,the number of active cores and hot-spot temperatures,and the impact of the number of frequency-scaled cores on hot-spot temperatures are investigated experimentally.(2)Statistic analysis for hot-spots selection of processorsThe hot-spot is the hottest functional unit of a processor.To analyze the relationship between the workload characteristics and the hot-spot selection of processors,a statistic method for analyzing the hot-spot selection of processors is proposed.First,based on the steady-state thermal model of multicore processors considering workload variations,a statistic model is proposed to analyze the hot-spots at the granularity of functional units.The probability that a functional unit is the hot-spot can be calculated given a subspace of workload characteristics.Second,based on the proposed statistic analysis model,statistic algorithms are presented to analyze the probabilities that each functional unit is the hot-spot and the impact of various instruction type on the probabilities of hot-spot selection.Finally,the relationships between the hot-spots selection and the workload characteristics are analyzed experimentally.(3)Analysis of minimum peak temperature of performance-constrained multicore processorsIn order to reduce the analysis runtime and analyze the relationship between the workload characteristics and the minimum peak temperature of processors,an analytical method is proposed to analyze the minimum peak temperature of multicore processors for the frame-based task model.First,the instructions per cycle(IPC)and the proportion of various instructions(PVI)are introduced into the analyzing model of the steady-state temperature of multicore processors.Second,for the frame-based task model,the minimum peak temperature model of multicore processors is proposed,and the minimum peak temperature is derived as the function of the voltage,the task period,the number of instructions,the PVI and the number of active cores.Finally,the proposed method is verified by the experiment.The relationships between the workload characteristics,the number of active cores and the minimum peak temperature are analyzed experimentally.(4)Analysis of maximum throughput of temperature-constrained multicore processorsIn order to improve the accuracy of the performance analysis of multicore processors and analyze the impact of the workload variations on the performance,in the case that the temperature is constrained,a method for analyzing the maximum throughput of multicore processors is proposed considering workload variations.First,the steady-state temperature is derived as the function of the PVI,the throughput,the voltage and the number of active cores.Second,the necessary condition for the maximum throughput of multicore processor is proved,and the model of the maximum throughput is established.Finally,the proposed method is verified by the experiment.The relationships between the PVI,the IPC,the number of active cores and the throughput are analyzed experimentally.(5)Probabilistic analysis of steady-state temperature and frequency of multicore processors considering workload variationTo perform the probabilistic analysis of the temperature and frequency of processor,the IPC is employed as the workload characteristic.In the case that the IPC follows the normal distribution,a probabilistic method is presented to analyze the temperature and the running frequency of multicore processors.First,at the micro-architecture level,the dynamic power of processors is modeled as the linear function of IPC and running frequency,and the leakage power of processor is approximated as the linear model of temperature.The error factors are introduced into the linear model of the dynamic power and the leakage power.Second,the micro-architecture-level hot-spot temperatures of both active cores and inactive cores are derived as the linear functions of IPCs of all active cores.It is inferred that the hot-spot temperatures of both active cores and inactive cores follow the normal probabilistic distribution,based on the assumption that IPCs of all active cores follow the same normal distribution.The probabilistic distribution of the set of frequencies is determined for the zero-slack scheduling policy.Finally,the probabilistic distributions of the steady-state temperature and the frequencies of multicore processor are analyzed experimentally.