Design Of Finite Element Systems For Parallel Computers
Goehlich, Ralph Dietmar (1989) Doctoral Thesis, Georgia Institute of Technology, Atlanta.
The presented work provides a synthesis of parallel processing and finite element methods (FEM) considering both vector and multiprocessing computer architectures. Parallel processing will provide effective computing speeds needed for future large-scale analyses. The key to achieving improved performance on this new class of computers is the development of software designs tailored to parallel computing. The focus of this work is on structural analysis. New parallel procedures and software designs have been developed for linear and nonlinear static FEM solutions.
Vectorized program designs for element stiffness computations and the generation of nonlinear nodal force vectors were investigated. Two conceptually different approaches were identified: single element and multiple element vectorization. Both methods can be used to increase the effectiveness of today's finite element systems but the multiple element method shows superior performance.
A completely revised software architecture for linear static analysis is proposed which consists of six parallel program modules. These parallel modules incorporate a double level storage technique on the basis of a shared data base and common access files.
Most of the applied parallel procedures are free of interprocess communication and processor load balancing is assured with self-scheduling loops. Two key modules, matrix decomposition and displacement solution, involve substantial synchronization. For these procedures an in-depth performance study was performed which included the implementation of the parallel matrix decomposition module in a production FEM system.
A new parallel program design is presented for nonlinear static analysis. The basic approach is similar as in the linear case and above multiprocessing modules are incorporated in the nonlinear solution sequence. Two additional modules are proposed for the parallel computation of nodal element forces and a parallel Quasi-Newton algorithm. The multiprocessing performance is evaluated.
A comparative study showed that the proposed design concepts are well suited for general-purpose FEM systems using moderately parallel shared memory computers possibly with vector capabilities on each processor.
Manuscript: order via UMI