Scalable parallel implementation for 3D semi-implicit hydrodynamic models of shallow waters

This work presents a parallel implementation for 3D semi-implicit hydrodynamic models of shallow waters that scales in low-cost clusters of computers. The scalability of semi-implicit hydrodynamic models is limited due to the need of all-to-one/one-to-all communications at each simulation time-step as it is here shown. These communications are avoided taking advantage of a nesting implementation, which resolves, in addition to the model at the original grid resolution (nested), a model with a lower grid resolution (parent). Nesting implementations are normally used to simulate both global and local processes with less memory and execution time by using as nested domain just the area where local processes occur while the parent model simulates the complete domain; but here, it is used to improve scalability. A two-level processing structure is proposed for the parallel implementation: pipeline plus domain-decomposition. The resulting parallel implementation with two-level structure scales with a slope near one.