Acceleration of finite element analysis by parallel processing

The finite element code LAGAMINE developed since 1982 has been adapted to numerous finite elements and constitutive laws. The present researches on micro-macro modelling increase the requirement of CPU time. The parallelization of software is often divided into two main approaches: the coarse grain approach and the fine grain approach. For Finite Element software, the coarse grain method is usually characterized by the use of the mesh partitioning. In the fine grain approach, there is no partition of the mesh. The parallel program is made by the parallelization of instructions issued of the nalysis of dependences between the instructions, and the loops. The efficiency of all these approaches depends on the organization of the memory and on the used exchange protocol (MPI, PVM, or OpenMP). Our approach is intermediate between these two main categories of parallelization. It is characterized by the parallelization of the assembly of the stiffness matrix using a coarse grain approach and the use of a parallel direct solver. Because of the architecture of the available parallel computers, which are shared memory, we decided to use OpenMP protocol.