Tests on parallel machines of a domain decomposition method for a composite three dimensional structural analysis problem

Abstract

We report about parallel implementation on LCAP1 system and on CRAY2 supercomputer of a parallel algorithm based on a domain decomposition method with Lagrange multiplier, for solving an ill conditioned three dimensional composite structural analysis problem, with as many as one million degrees of freedom. We show that this method has very good features on both granularity and data dependancy viewpoints. The tests performed prove that the choice of the local solver is very important to get an efficient global method. For the studied case, it is clear that solving the local problems with a direct method is the best solution, and we give some reasons why it will be the same for many other problems.