Research On Performance And Energy Consumption Co-optimization Of OpenMP Program In Power Constrained HPC System

With the increasing demand for computing power in scientific research and production and life,high-performance computing systems have rapidly developed in the direction of E-level tens of billions of calculations,but energy consumption has become an important obstacle to the development of high-performance computing systems.This puts forward higher requirements on the architecture design of the hardware system and the execution efficiency of the software system.In future high-performance computing systems,power constraints will become a commonly used energy-saving means,and different computing nodes and computing tasks will operate according to the given power ceiling.As a commonly used programming model for parallel computing,OpenMP parallel programs have been widely used in different computing scenarios.Studying the operation of OpenMP programs under power constraints can help promote the development of high-performance computing systems.This paper proposes a strategy for fine-grained performance and energy consumption collaborative optimization of OpenMP parallel programs based on a power constraint system.This strategy achieves the goal of power constraints by setting different CPU and DRAM power caps on the parallel domain of the program,and at the same time Machine learning modeling of energy consumption can predict the performance and energy consumption of program execution under different power upper limit settings according to actual needs.Based on this strategy,this paper designs and implements an automatic code optimization system,which can select the optimal power configuration according to different optimization goals,collaboratively optimize the performance and energy consumption,and modify the optimization program at the source code level.In this paper,twelve sets of OpenMP benchmark test programs were executed on the experimental platform for training and prediction,and the optimization effect of the strategy we proposed was verified.Compared with the original program,in the case of optimizing energy consumption,energy consumption can be reduced by an average of 9.25%,up to 18.75%,while in the case of the optimal energy delay product,the energy consumption can be reduced by an average of 4.74% without greatly affecting performance,and a power reduction of 30.4% can be obtained.Experimental results show that the optimization strategy we proposed can optimize the power configuration,performance and energy consumption of the OpenMP program under power constraints.