Manufacturing of weather forecasting simulations on high performance infrastructures

The Weather Research and Forecasting (WRF) model represents a mesoscale numerical weather prediction system designed for both the atmospheric research and operational forecasting use. WRF offers several computationally efficient platforms while providing serial and parallel models running with or without multi-threading. The main goal of the work described in this paper was to develop management software tools facilitating the execution of WRF simulations on High Performance Computing (HPC) infrastructures. The management tools enable easily to carry out the complex simulation process, representing a workflow of central jobs, i.e. preprocessing, modeling and post-processing, which are running primarily on our local HPC cluster with the possibility to exploit also the computational power of the European Grid infrastructure. The second objective of this work was the investigation of parallel MPI and hybrid parallel MPI+OpenMP WRF models using the created tools to run simulations. The aim was to find out the most suitable combination of hardware and software configuration setting for the realization of the WRF simulation process with the given input scenario. Performance results of real simulation processes running on various hardware configurations of the cluster are presented.

1. Introduction

In recent years, along with the rapid development of information technologies and extensive amounts of data observed also the field of mathematical modeling and computer simulation has enjoyed intense advances in both directions of data and computational intensive strategies. Computer simulations are of great importance mainly in such situations when it is impractical, financially expensive or too risky for the realworld system to be directly subjected to experimentation, or in case of evaluating a system before it is actually built [1], [2]. Applying simulations, the behavior of the event can be predicted based on a set of parameters and initial conditions, and solutions can be found out before time, money, and materials are invested. In most cases simulations with a real scenario represent a long-time, computational intensive and memory consuming job. Moreover, both the model calibration and also the research objective itself calls for performing a great number of simulation runs using diverse input scenarios with variable input parameters and boundary conditions. For these facts it is unavoidable for the researcher to have a HPC infrastructure in hand which is capable to accomplish its task taking time as short as possible. The second no minor important issue is understanding of the middleware necessary to execute applications. For non-informatics scientists it is often too complicated to be used correctly. A promising approach has proven to employ various application supporting instruments that make easier the process of the submission and execution of simulations. Hiding much of the middleware complexity they can insulate the researcher from the computing platform and give him the impression that all of the used resources are available in a coherent virtual computer center.