Massively Parallel Multifrontal Methods for Finite Element Analysis on MIMD Computer Systems

The development of highly parallel direct solvers for large, sparse linear systems of equations (e.g. for finite element or finite difference models) is lagging behind progress in parallel direct solvers for dense matrices and iterative methods for sparse matrices. We describe a massively parallel (MP) multifrontal solver for the direct solution of large sparse linear systems, such as those routinely encountered in finite element structural analysis, in an effort to address concerns about the viability of scalable, MP direct methods for sparse systems and enhance the software base for MP applications. Performance results are presented and future directions are outlined for research and development efforts in parallel multifrontal and related solvers. In particular, parallel efficiencies of 25% on 1024 nCUBE 2 nodes and 36% on 64 Intel iPSC860 nodes have been demonstrated, and parallel efficiencies of 60-85% are expected when a severe load imbalance is overcome by static mapping and dynamic load balance techniques previously developed for other parallel solvers and application codes.