Parallel Finite–Element Analysis via Message Passing

Abstract

This paper describes the evolution of a finite–element program with explicit time integration from a sequential form to a parallel message passing incarnation. The resulting parallel code has been used on a wide range of parallel platforms. The finite elements are allocated among multiprocessors through various domain decomposition strategies. Separate input and output files are established for each subdomain. These files are read from and written to by local copies of the program executable operating in parallel. At each time step, only internal forces for nodes at subdomain interfaces are exchanged. The Argonne–developed p4 package is used to manage interprocessor communication. Program fragments are presented to illustrate use of p4 in the context of finite–element analysis with explicit time integration. The code is portable among platforms supporting Berkeley UNIX 4.3BSD interprocess communication sockets. Results are presented for problems exhibiting large–deformation, elastic–plastic dynamic transient response. Benchmarks highlighting the benefits of parallel computing on both the small and medium–sized problems that engineers run routinely as well as large–scale ones are included. Speed–up factors are presented for shared–memory MIMD multiprocessors, networks of workstations, and two distributed memory MIMD multiprocessors—an Intel i860 Hypercube and the Intel Touchstone Delta supercomputer.