Research On Key Technologies Of Energy Efficiency Optimization Of Compute Nodes In Parallel Computer System

With the development of supercomputers,high energy consumption has become a important factor affecting the further development of supercomputer.The high energy consumption has become the main cost in supercomputer center.At the same time,high power can bring system stability and heat dissipation problems to the supercomputer system.With the improvement of performance,it is necessary to consider energy issues to improve energy utilization and obtain overall optimization of energy consumption and performance.Compute nodes play an important role in performance,power consumption and energy consumption,so it is very important to improve the energy efficiency of compute nodes.Today,there are a lot of researches on the energy efficiency of compute nodes.But these researches do not take into account the overclocking of the processor,so that they can not get the optimal energy efficiency in some cases.In addition,the impact of the mapping relation between processor cores and processes in multiple compute nodes is not considered.Based on these problems,the research content of this paper is mainly from the following aspects:Firstly,this paper studies the impact of overclocking on energy efficiency of different applications in real hardware platform.We use Energy Delay Product(EDP)as the evaluation criterion for energy efficiency.The lower EDP means the better energy efficiency.HEE algorithm is proposed.According to the characters of application and real hardware,HEE algorithm gives a scheme to improve energy efficiency.The experimental results show that HEE algorithm can obtain an average 13% energy efficiency improvement,and the maximal energy efficiency increase is up to 25.6%.Secondly,this paper give a way to improve performance without the increase of power consumption in single compute node.The CEE algorithm is proposed,which divides the application based on the characteristics of the program.For memory-intense applications,use memory overclocking to improve performance and control power consumption by scaling down processors frequency.For applications with less memory access,use processor overclocking to improve performance and control power consumption by turning off some memory channels.The experimental results show that CEE algorithm can obtain an average of 9.3% performance improvement.Meanwhile the total power consumption does not increase.The maximal performance improvement is up to 13.1%.Thirdly,this paper achieves the goal of improving the performance without the increase of power consumption in multiple nodes.In a supercomputer center,a task usually needs use multiple nodes at the same time,and the number of process is decided by user.In this paper,the TPM algorithm is proposed.By analyzing the application characteristics,the algorithm achieves the goal of improving performance without increasing the peak power by reasonably mapping the processes to the processors core and adjusting the processors frequency.The TPM algorithm can achieve an average 12% performance improvement,and the maximal performance improvement is up to 20.2%,while ensuring that peak power is not increased.