Abstract
Numerical Weather Prediction (NWP) requires real-time, high-accuracy processing, straining traditional high-performance computing clusters with limited resources, complex operations, and long queue times. Hybrid clouds merge the security of local clusters with the scalability of public clouds, providing a viable solution for high-performance computations. However, it also poses challenges: parallel programming for local clusters is not suitable for the various settings of hybrid clouds; complex parallelization policies increase communication overhead and complicate scheduling; and traditional static resource binding can lead to load imbalance in heterogeneous environments. This paper proposes an adaptive cross-platform scheduling strategy tailored to the characteristics of NWP models. This approach harmonizes the advantages of traditional and cloud-based parallel computing, integrating two distinct parallel programming methodologies and reconfiguring the parallel programming framework of the forecasting models. Experimental results show that the framework effectively improves adaptability and resource utilization, significantly improves computational efficiency and reduces operational overhead in hybrid cloud deployments.