DISTRIBUTED ALGEBRAIC MULTIGRID FOR FINITE-ELEMENT COMPUTATIONS

The Finite Element Method has been successfully applied to a variety o f problems in engineering, medicine, biology, and physics. However, th is method can be computationally intensive, particularly for problems in which an unstructured mesh of elements is generated; In such situat ions, the Algebraic Multigrid (AMG) can prove to be a robust method fo r solving the discretized linear systems that emerge from the problem. Unfortunately AMG requires a large amount of storage (thus causing sw apping on most sequential machines), and typically converges slowly We show that distributing the algorithm across a cluster of workstations can help alleviate these problems. The distributed algorithm is run o n a number of geomechanics problems that are solved using finite eleme nts. The results show that distributed processing is extremely useful in maintaining the performance of the AMG algorithm with increasing pr oblem size, particularly by reducing the amount of disk swapping requi red. (C) 1998 Elsevier Science Ltd. All rights reserved.