9625865 Szyld The investigator studies block and asynchronous iterative methods for the efficient solution of nonsymmetric linear algebraic systems of equations. One aspect of the project is the study of the application of threshold partitioning algorithms for sparse matrices. The emphasis of the project is on singular systems. Singular systems of equations arise in many applications, including certain types of differential equations modeling physical phenomena, Markov chains, and queuing models. Markov chains are used, e.g., to evaluate the performance of computer systems, while queuing models in particular simulate traffic in telecommunication networks. This project develops fast methods for solving linear systems of equations, particularly singular systems. The methods are iterative: starting from an initial guess at the solution of a system, the method produces a sequence of approximate solutions that approaches the actual solution. Asynchronous methods for parallel computers refer to methods in which part of the computation is performed in each processor in an asynchronous way, i.e., without synchronization between the processors. In other words, each processor proceeds with its tasks, using the information available at any given moment without waiting for the other processors to finish their tasks. Processors do not wait: they keep active computing better approximations to the solutions. Therefore the overall solution time can be greatly reduced. The variables of the differential equations, of the Markov chains, or of the queuing models, may be partitioned naturally into sets of variables, e.g., by the near-connectivity of the nodes of a network or graph. This partition induces a particular block structure in the matrix representing the linear system to be solved. Asynchronous parallel methods are well suited to deal with these blocks. New partitioning algorithms that take into account the values of the matrix are explored. The discre tizations of many differential equations, such as certain convection-diffusion equations, give rise to matrices with coefficients varying over a wide range of values, i.e., some large and some very small. These problems are very hard to solve in practice, and most modern iterative methods fail to converge. The methods that do converge are very expensive or very slow. Partitioning algorithms that take into account these values can be used to permute these matrices so that the large coefficients lie along the diagonal of the matrix. Parts of this permuted matrix, e.g., its blocks along the diagonal, can be used as an effective preconditioner. Preliminary experiments indicate that this approach leads to fast convergence. Thus, the project can provide a method to efficiently solve a class of important problems in science and engineering.
