Several general purpose computer systems with multiple processors operating concurrently are currently being commercially produced. Most of the present generation of finite element software was not designed to take advantage of this new technology. The purpose of this paper is to present the advantages of a new architecture for finite element programs which will operate efficiently on computer systems with any number of processors. Also, the basic approach is effective for multiprocessor computers with local or shared memory.
The new computer program architecture is based on an initial application of a simple algorithm which automatically subdivides the complete finite element domain into a number of subdomains equal to the number of available processors. The resulting data structure requires a minimum of communication between processors during the formation of basic element matrices, reduction of subdomain matrices and the postprocessing of element results. The assembly and solution of the global system of subdomains can be accomplished directly or iteratively, and new concurrent solution algorithms can be introduced.
Several of the ideas presented here have been tested on various types of computer. The new program architecture indicates that it is possible to obtain speed-up times of over 90 per cent of the maximum theoretical values if appropriate numerical methods are employed.