| The earth system model covering physics,chemistry,biological activity,computer and many other disciplines is an important numerical simulation tool used by researchers in the climate field to study and understand the mechanism of climate and environmental evolution and to predict future global climate change.And cause of the earth system model is of great significance in studying and predicting the global climate change,the mainstream countries in the world have already carried out relevant research and development work.The most representative of many research results is the coupled climate model CESM(Community Earth System Model)developed by NCAR.The coupled climate model CESM uses a pluggable modular coupling frame structure,which is the most advanced open source coupled climate model in the earth system model.It also has a huge impact on the coupled climate system model developed by many other countries and laboratories.According to statistics,in CMIP5 more than a quarter of the coupled climate models are directly related to the coupled model CESM or use its underlying framework,which also includes China's climate coupled climate models submitted by the Model Development Lab.The Earth system model software used by researchers is mainly deployed in high-performance clusters or supercomputers.The coupled earth system mode,such as the coupled model CESM,supports multi-process,multi-thread,and shared memory,and its internal also has a complex coupling relationship between each component model,thus the communication,memory access and calculation has the characteristics of relatively dense,and these factor contain each other affects the performance of the whole model software..At present,there is still much work to be done to optimize the operational performance of the climate mode represented by CESM.According to research statistics,every 1% increase in operating performance of the CESM will save researchers millions of dollars in expenses.In addition,the performance optimization work for the CESM can be directly or indirectly applied to other coupled climate models of the same framework,which is also of great significance for the research and development of coupled climate models currently being carried out in China.In order to optimize the operational performance of the coupled climate model CESM,this article first analyzes the main structure,operation and communication behavior of the CESM on the source code,and then combines the actual operation statistics of the CESM to analyze the CESM.And analysis of the operational behavior of the component model.The analysis shows that each component model in the coupling mode has different performance under different component settings,and the coupling mode itself has fluctuations in operating performance with or without external interference.After completing the related analysis work on CESM,this paper models and analyzes the resource mapping allocation problem of each component mode in the coupled mode,and abstracts this process into a special two-dimensional binning problem.In order to solve such a problem of resource mapping allocation,this paper proposes an optimal resource allocation algorithm GSLOpt based on historical statistics.In order to solve the possible impact of the coupling model performance fluctuation phenomenon,this paper proposes a set of online resource adjustment strategies for the coupling model.Based on the analysis of fluctuations and historical data in the strategy,an AHBOS algorithm strategy for revising historical statistical data is proposed.Due to the need for online resource adjustment to obtain a redistribution plan in time,this paper proposes a highly efficient heuristic algorithm FFADH for online resource adjustment.Finally,the proposed resource mapping allocation algorithm and online resource adjustment strategy are integrated to implement a set of tools for automatic resource allocation and online adjustment for the coupled model.In the experimental analysis,the resource mapping allocation schemes of the proposed optimal resource algorithm GSLOpt and the FFADH algorithm for online resource adjustment are applied to the instances of the coupling model respectively.A comparative analysis of examples of common resource strategy programs.It is found through experiments that the instances of the GSLOpt algorithm scheme have the best performance in the experiments under the same component settings and mesh division.Compared with other strategies in the experiment,the average iteration time of the instance using the GSLOpt scheme is optimized by at least 9.9%,and the simulated years / cmp-day is improved by at least 9.07%.The FFADH algorithm for online resource allocation has good time complexity.In the experiment to verify the FFADH algorithm,,the instance using the FFADH algorithm scheme has an average iteration time optimization of at least 4.1% in operation performance and a minimum of 5.22% improvement in simulated years / cmp-day.For the experimental analysis of the online resource adjustment strategy,this article will use an instance of online resource adjustment and a normal instance to run in a cluster that simulates load changes in reality.Compared with the normal instance,the instance using online adjustment of resources in the experiment can achieve an average improvement of 8.97% in overall running time.In the extreme load experiment,using the example of online resource adjustment,the average iteration time after the resource adjustment can be reduced by up to 12.28%,and the simulated year / day can be increased by up to 9.36%. |
PDF Full Text Request