High parallelization efficiency in barotropic-mode computation of ocean models based on multi-grid boundary ghost area

In general, barotropic-mode computation requires the largest communication time in ocean models because of its iterative nature, when parallel computation is performed based on regional partitioning. In this study, reduction in the overhead time included in the communication in the parallel computation of the barotropic mode is considered to achieve a high parallelization efficiency of ocean models. We verify that the reduction in the communication frequency based on the multi-grid boundary ghost area reduces the total communication time. We find that this is because the sum of several kinds of overhead time for communication occupies a fairly large part of the total communication time. We discuss the trade-off between the decrease in communication time and the increase in computation time due to increased boundary area in such cases, leading to a determination of a suitable width of the multi-grid boundary that minimizes the total required time. We also discuss the efficiencies of the one- and two-dimensional partitioning of the model domain, when a multi-grid boundary is used. In general, two-dimensional partitioning is more efficient than one-dimensional for large partitioning numbers.