On Parallelism And Scalability Of An IterativeModel For Coastal Seas Fluid Dynamics

The efficiency and the scalability of a successful parallel implementation of a 3D transport model for coastal seas is presented. The work is concentrated on the evaluation of the parallelism of the iterative solver after implementing several High Performance Computing techniques. Thus, the impact of a block variant of the ELU preconditioner proposed to avoid the inherent sequentiality of this step is measured. This variant was obtained by a spectral bisection technique which is used to partition the FEM mesh. Consequently, each submesh corresponds to a submatrix (a block) which can be treated in parallel. The main impact is that the number of the neglected coefficientes is minimised, and thus, the convergence degradation is not as severe as would be produced by a block ELU version without partitioning. In order to accelerate the convergence of the BiCGSTAB solver even more, a matrix ordering algorithm is applied for each submesh locally. Therefore, a high ratio of scalability and a good efficiency is achieved. The MPI protocol was used for this parallel implementation on a IBM SP2 machine. The model performance in terms of efficiency and scalability is evaluated with respect to the problem size and the number of processors.