Study On Linear Finite Element Method In Cloud Computing

Cloud computing that based on virtualization technology is now the focus ofresearchers, business models and academic researches concerning it appears continuously.In the past few decades, high performance computer with multi-core and multi-processorhas provided an effective solution for large scale complicated engineering and scientificproblems, however, the high performance computing applications are quite professionalwith rather high threshold. MapReduce is a common used programming model in cloudcomputing, it’s a highly abstract model for parallel descriptions. With it’s simpleprogramming interface, the program can run auto-parallelization, this greatly reduces thedifficulty in parallel computing program development. Finite element method (FEM) hasgreater demand for computing resources than before, as the scale of engineering problemsis increasing all the time. Its implementation based on cloud computing can be used tosolve more complicated engineering problem with less computing resources and it is arather important exploration.In this paper we made a detailed investigation on the migration of scientific computingto cloud computing, then we built a cloud computing environment by Hadoop, andpracticed the MapReduce programming model in numerical computing. The parallelstrategy and implementation of the sparse matrix vector product (SPMV), conjugategradient algorithm (CG) and other numerical method were also investigated in this paper.We analyze the characteristics of FEM and the advantages of MapReduce programmingmodel, and then migrate the FEM code to Hadoop platform by an appropriate datastructure, this include the element stiffness computation and system stiffness generation.The strategy and algorithm for spatial truss linear system calculation runningauto-parallelization were presented in this paper.Numerical results show that, it is simple and easy to use the proposed strategy and MapReduce model to implement the FEM calculation in Hadoop. This paper achieved agood performance in SPMV and the FEM assembly step, and we find that the efficiencybecomes better with the increase of the size of the problem as well as the number ofcluster’s node, the efficiency in solving FEM system is limited but can be improved in thefuture.