Lessons learned from attempts to parallelize linear FE analysis code for 2D structures

This paper discusses the lessons the author learned in his attempts to parallelize linear finite element analysis (FEA) code for 2D structures using MPI. Three versions of an FEA code obtained by using three different solution techniques were parallelized using the method of domain decomposition. The parallel programs showed excellent speed up on an IBM SP1 when 2 and 4 processors were used. The speed up was not quite as good when 8 or 16 processors were used. It is suspected that the communication pattern to assemble and disseminate nodal information is the bottleneck when more processes are used. The future course of work on the subject is discussed