MIMD implementations of linear solvers for oil reservoir simulation (Q2638722): Difference between revisions

The paper presents a set of numerical experiments to solve linear systems arising from oil reservoir simulation, because solving them represents half of the overall computational effort involved in the simulation. The authors compare several basic iterative algorithms that they have implemented on a transputer network: conjugate gradients, Jacobi, Gauss- Seidel and SOR. They also discuss decomposition strategies for each method. Their conclusion is that conjugate gradients are a far better choice than any other; second comes the Jacobi method. A discussion of the implementation of incomplete Cholesky conjugate gradient follows and the authors find it relatively disappointing, because the performance worsens with respect to plain conjugate gradient, as the number of processors increases; they reasonably attribute it to the fact that the preconditioning step (the Cholesky decomposition) is of sequential nature. The authors finally propose to shift to the parallel Jacobi iteration as a preconditioner for conjugate gradients and the numerical experiments show a significant improvement over the conjugate gradient method.