Energy Efficiency Optimization On Chip Multiprocessors Under Power Limitation

Modem processors are equipped with power regulation mechanisms,so it could adjust processor power to meet different scenarios and requirements.Power capping is one of the most important applications.Controlling power means controlling temperature,which could improve reliability.In data centers,power oversupply and infrastructure investments could be cut down by capping cluster power.Moreover when it is necessary to add more servers into clusters,it's safe to not modify current power supply facility.With the increasing number of processor cores and more complex of runtime workloads,power controlling algorithms are facing the problems of too complex and lacking of practicality.Most optimization strategies depend on offline training,which may damage optimization results and portability.This paper proposed PPCM(Practical Power Capping Method):a multi-level power limitation and optimization method.PPCM is stable and has strong robustness by utilizing dynamic linear model and feedback control theory.PPCM uses the ratio of computation to memory access as an indicator to allocate power among applications,then considers the features of multithread programs and allocates power among threads which are in the same program.It improves energy efficiency by allocating power to processor cores which need more power.The main research and achievements in this paper are listed as follows:1)This work analyzed the impacts of three power regulation methods on processors when running workloads,including Dynamic Voltage and Frequency Scaling,Idle Cycle Injection and Clock Cycle Modulation.First,we analyzed the impacts on CPU power,energy and performance.Then we contrasted these three power regulation methods,discussed their similarities and differences,as well as applicable scenarios.2)This work designed a dynamic adjustment method' which is based on feedback control theory.After studying the relationship between processor power and frequency,we use linear model to estimate processor power.Then utilizing P-control method to cap processor power.The scale factor in linear model varies with workloads and hardware.So we designed a dynamic adjusted power control method.3)This work proposed a multi-level power allocation strategy.It has two levels:application-level power allocation and thread-level power allocation.On the first level,it utilizes the ratio of computation to memory access as an indicator to allocate power among applications.On the second level,it allocates power among threads by reference to memory access.At the same time,we took into consideration that frequency change memory access rate.So,we normalized the memory access overhead on different frequency to the same frequency.Experiments show that,PPCM can control CPU power effectively and accurately.The error between CPU actual power and power budget is within 4%.Compared with Priority algorithm,PPCM improves performance by 10.7%,saves energy by 5.1%on average,and EDP is reduced by 14.3%.PCMCA is an outstanding algorithm at present.PPCM is superior to PCMCA by 4.5%in performance and 5.0%in EDP.