Multilevel Parallel Solution of Large, Sparse Finite Element Equations from Structural Analysis

We discuss the use of high performance computing equipment in large, commercial structural analysi s programs. In particular we consider the strategy for modifying a standard industrial code from a pure F77 version to a form suitable for a range of parallel computers 4. The code is parallelized using the PVM message passing library for communication between processes and the ScaLAPACK and BLACS libraries [4, 7, 8] for parallel linear algebra operations. The parallelized code is suitable for a range of parallel computers, however for the purposes of verification and benchmarking, two specific hardware architectures were targeted in this work. These are an 8-node DEC-Alpha cluster with 233MHz EV45 processors and FDDI/GIGASwitch interconnect, and a 32-node Parsytec GC/PowerPlus with 64 PowerPC-601 processors and Transputer interconnect.