The ocean which covers the majority of the surface area of the earth has always been the focus of scientific research. It has been a long time since ocean simulation started using high-performance computers. Currently, not only various wind wave models and ocean current models are developed, but there are some couple models, which combine elements like atmosphere, ocean, sea ice, etc. With the increase of computational accuracy, the requirements for computational resources, memory bandwidth and network latency are also largely increased
Different applications in ocean simulation usually have different characteristics. We analyze these characteristics by taking the ocean numerical model FVCOM and global ocean model NEMO as an example. ●When in operation (as shown in the left figure), FVCOM needs a small memory space and a moderate memory bandwidth, but inter-node continuous data streams communication bandwidth is large. This indicates that a high-bandwidth and low-latency network can ensure good computational efficiency when FVCOM is running. Additionally, no special requirement is presented to I/O because the amount of data written to the disk is not large, although writing is continuous during computation. ●When in operation, NEMO (as shown in the right figure below) requires a large memory space and also a large memory bandwidth. Therefore, a large memory space needs to be allocated to each computational node and the need for memory bandwidth must be met. In contrast to FVCOM, when NEMO is running, amount of data communication between nodes is small and is increased only during periodic data write operations, hence no special requirement is presented to network bandwidth.