Real-time Power Estimation For Complete System On Multi-core Platform

High performance makes development of many categories of processors.However,according to the Moore's Law,the power consumption of processors has increased exponentially,thus affecting system performance and accelerating aging of the chip.Power consumption has aroused many people's attention.Estimating real-time power consumption of systems has been pivotal for hardware and software system research and design.This paper aims to constructs two kinds of power models to estimate the power consumption.We analyzed the activities of the processor,solid state disk,mechanicak hard disk and memory,such as the CPU frequency,each core's utilization and performance events and so on.The contributions of the thesis are as follows:1.We proposed system-level power models for a DBMS server.These models consider the CPU frequency and the utilization of each cores,disk and memory.These models have good portability and accuracy.To improve the accuracy of power models,we investigate various combinations of components and power factors in models.Meanwhile,we investigate the effects of solid-state disk and mechanical hard disk on the models.2.We constructed another kind of power models that use the information of performance counters.A feature selection procedure was performed by analyzing the relationship between performance events and power.3.We constructed real-time power models for full-system based on performance events and device utilization.The models integrate the static power,the dynamic power of processor,memory and disk.They are independent of the running workloads.Experiment results demonstrated that the models designed for the DBMS server has a good portability and accuracy.And the relative errors of models based on utilization are between 5%and 8%.The relative errors of models designed for a processonr and memory are between 9%and 13%.The relative errors of models based on performance events and utilization is between 4%to 13%.