Using the technique of Message Passing Interface, a parallelized version of a coupled wave-circulation model was set up. The tested model is a regional one for simulating the seas off China, which is comprised of 450,625 elements and 30 vertical sigma layers. The implementation efficiency was evaluated on two kinds of computers, the HP Integrity Superdome and SGI Altix 4700 multiprocessor. The numerical results show that the low-communication high-extra-computation scheme (LCHC) produces higher efficiency than the high-communication no-extra-computation scheme (HCNC) while the number of processors exceeds 24 for HP Integrity Superdome and eight for SGI Altix 4700, respectively. The experiments with both LCHC and HCNC scheme show super-linear speed-up when the number of processors is small. The model with the LCHC scheme is preferred as it achieves parallel efficiency in excess of 90% on the HP machines for all experiments with the number of processors no more than 100, while the efficiency decreases rapidly with the HCNC scheme after the number of processors increases to more than 24. Numerical results suggest that the parallelization of this coupled wave-circulation model is efficient and portable to a variety of parallel architectures.
Read full abstract