Design And Implementation Of Parallel Framework For Large-scale Hydrological Simulation Calibration

In hydrological research,hydrological models are mainly used in the field of water cycle research,and they are beneficial tools for scientific evaluation and rational management of water resources.These hydrological models have many parameters with clear physical meanings and significant spatial-temporal distribution characteristics,so the calibration of parameters is particularly important in the entire hydrological simulation.With the emergence and development of supercomputers,high-performance parallel computing has been successfully applied to hydrological models.However,the current parallel methods cannot achieve optimization of large-scale hydrological simulation parameters.The main reason is that most of the parameter calibrations use the master-slave parallel mode of blocking communication.In the process of large-scale parameter calibration,it is difficult to achieve load balancing between processes,and the parallel efficiency is far lower than the ideal value.In view of the above problems,the main research contents of this dissertation are as follows:(1)A multi-level parallel calibration framework is designed and implemented.The framework adopts the peer-to-peer design mode,non-blocking communication method and sub-communication domain technology in the MPI parallel library.The peer-to-peer design mode enables each process to participate in parameter calibration and maximize the use of resources.The non-blocking communication method can ensure all communication operations to run in the background,and avoid the inter-process waiting caused by communication.The sub-communication domain technology realizes the multi-level parallelism of the framework,and divides multiple processes into different communication domains.Each process has its own division of labor and accomplish parameter optimization together while calibrating parameters.(2)The combination of optimization algorithm,hydrological model and multi-level parallel calibration framework is realized.According to the characteristics of particle swarm optimization(PSO),part of the processes in the parallel framework can be mapped to search particle points,and part of the sub-communication domains can be mapped to populations,so that more sub-communication domains can approach the optimal value,so that the whole framework can converge more quickly,to achieve better calibration results.The optimization algorithms are applied to the HIMS model and the Easy DHM model.The experimental results show that the genetic algorithm and the particle swarm algorithm can exert an excellent calibration effect in the parameter calibration of the hydrological model.(3)The performance of the multi-level parallel framework designed in this dissertation is tested.By comparing the two-level master-slave parallel framework with the multi-level peer-to-peer parallel framework in three aspects: optimization effect,running time and scalability,it can be verified that the peer-to-peer parallel framework designed in this dissertation has better effect in large-scale parameter calibration.The parallel framework is tested by increasing the scale of hydrological simulation and the number of processes,and the final result verifies that the parallel framework has good scalability.