Author(s): Ang Zhang; Tiejian Li; Xiang Li; Jiahua Wei; Liguo Zhang; Haiyun Shi
Linked Author(s): Xiang Li, Jiahua Wei
Keywords: Parameter calibration; Genetic algorithm; HPC job scheduling; DYRIM; Parallel computing
Abstract: Numerical simulation models have been one of the most important tools for analyzing real-world problems. Undoubtedly, the increasing requirement of numerical computations raised the computing time problem. Running a parallel model in a multi-processor computer or a cluster of computers can help overcome the single processor limitation and has been widely implemented in recent years. Even so, computing time is still an impediment when an optimization technique is applied to model parameter calibration. Trial-and-error is a traditional way to determine the model parameters. However, such technique requires considerable experience of sophisticated modelers or researchers, and would thus be inapplicable to common engineers or with new models. It is in this context that we propose a general framework for model parameter calibration using job scheduling, which is one of the key capacities of a high-performance computing (HPC) system, e. g., the Windows HPC Server used in this paper. This methodology can be used for various models and optimization techniques. In this study, the model to be calibrated is specified as the Digital Yellow River Integrated Model (DYRIM), a river basin model developed in Tsinghua University. To parallelize model simulation, the DYRIM implements a dynamic spatial domain decomposition method with the Message Passing Interface (MPI) running in HPC system. The genetic algorithm (GA) is used as the optimization technique and developed with the HPC job scheduling interface in the proposed framework. When GA needs to evaluate the fitness of various model parameter combinations, the job scheduler of the HPC system is called to allocate computing resources to a number of jobs simultaneously execute the model with the parameter combinations. After the jobs’ execution, the GA collects the model efficiency estimations and proposes a new generation of parameter combinations, until the stop criterion is reached. The specified model and optimization technique are applied to the Qingjian River basin in the Middle Yellow River in northern China. Results demonstrate the framework can effectively make use of computing resources and significantly reducing the computing time, as well as explore the model behavior under different parameter combinations.
Year: 2013