On the parallelization of unsteady BEM problems with variable mesh size

This paper presents code acceleration strategies for a boundary element method (BEM) simulation of an unsteady free-surface problem with a variable mesh size. The evolution of the film and droplet field emanating from a classical swirl injector/nozzle is used as a test case for the methodology. The unsteady free-surface problem (dynamic grid problem) presents challenges for load balancing the inversion of a matrix that is changing size with time. Since inversion of the matrix of linear equation coefficients represents the dominant computational cost for large problems, the parallelization of the second-order BEM code is implemented using a portable computation library, ScaLAPACK. A different processor grid topology using ScaLAPACK has been implemented and compared to the recommendation of ScaLAPACK guide. Implementation of ScaLAPACK in the solver routine of BEM has been successful and a significant simulation time reduction is demonstrated on an Opteron Dual Processor high-performance computing cluster employing two processors per node.