| With the continuous development of mechanical industry and computer science,numerical simulation is widely used in every stage of mechanical design and manufacturing,which greatly reduces the cost of experiment.In the field of numerical analysis and industrial simulation,the finite element method is one of the most widely used numerical calculation methods.Traditional numerical analysis programs run on a single computer or server can easily reach performance bottlenecks,slow down the efficiency of calculation and even fail to complete relevant calculations when faced with large-scale complex examples.In recent years,the cluster distributed parallel computing gradually rises with the wide application of big data,and the cluster platform has excellent computing power scalability in theory,so the cluster distributed framework has also become the key technology of cloud computing.In this paper,we study and develop a cluster of distributed parallel finite element method computing platform,based on the memory is used to calculate the Spark of cluster distributed framework,its operation in the virtual machine,the hardware platform of the JVM has extremely high the movable property,using the framework of cluster resources of abstraction,developers can focus more on algorithm of research and development,the development and the maintenance efficiency is greatly increased.Combined with the finite element theory and high performance programming idea,this paper has done the following related work based on Spark cluster distributed framework:(1)Based on Spark framework,according to the finite element theory and solution steps,the data flow and overall architecture of the finite element method parallel computing platform are studied.Following the development principle of aiming at high performance and iteration,the overall framework of each calculation module of the finite element with high decoupling is designed by using the distributed programming idea.It serves the specific development of the module and the modification and addition of the algorithm of the computing platform in the later stage.(2)Design and develop the INP file reading algorithm based on multi-threading,stiffness matrix assembly algorithm based on distributed computing,boundary condition application algorithm based on sparse matrix,linear equations solving algorithm based on distributed computing and VTK project file output algorithm.The design of solving module for linear equations is discussed in detail.The distributed Gaussian elimination method and distributed conjugate gradient method based on Spark are developed by making full use of the performance and resource advantages of distributed computing.In this paper,design and development of the finite element method distributed parallel computing platform in the laboratory of 3 host cluster environment,through the numerical experiments show that compared to the stand-alone version on the computing capacity and computing efficiency all have certain advantages.The gaussian elimination method has more than 5.327 times of calculate capacity,speedup ratio is2.098.The conjugate gradient method has more than 18.078 times of calculate capacity,speedup ratio is 166.135.At the same time,the algorithm has a high expansion ratio and acceleration ratio,and the error of the experimental results is very small compared with the calculation results of ABAQUS commercial software,so the algorithm has a very high computational accuracy.The successful design and development of the platform has made another exploration for the finite element method cluster distributed parallel computing research,and has left valuable experience for the development of scientific computing algorithms based on the JVM and other general platforms.At the same time,it has made an attempt for the finite element method in cloud computing. |
PDF Full Text Request