Parallel efficiency in implicit multiblock, multigrid simulations, with application to submarine maneuvering

Abstract

Large-scale complex flow simulations require substantial parallel computing resources. The present paper addresses some considerations of parallel efficiency in connection with an iterative implicit multiblock, multigrid algorithm for unsteady incompressible viscous flows. The algorithm is an implicit finite-volume Roe/ MUSCL spatial approximation, combined with discrete state-variable flux linearization, nonlinear multigrid iteration at each time step, and scalable concurrency introduced by a block-Jacobi LU/SGS iteration at each multigrid level. Semi-empirical performance estimates for parallel CPU, memory and cost efficiencies are described for an MPI implementation of this algorithm on existing and hypothetical computing platforms. These estimates are used to develop scalability and efficiency predictions in the form of performance landscapes for both memory-constrained sizeup and constant-problem-size scaleup modes. The estimates include parameters such as MPI software bandwidth and architecturespecific software tuning. These results indicate that the method is scalable in a practical sense for large-scale problems. Recently computed results illustrating a large-scale simulation for a submarine maneuver are given, specifically, a rising maneuver induced by a prescribed motion of sailplane appendages.