An optimization framework for adjoint-based climate simulations: A case study of the Zebiak-Cane model

摘要

Conventional adjoint-based climate simulations are generally time-consuming because they require the frequent reading of historical data from hard disks, adopt a calculate-all implementation strategy, and work in a serial way. To improve the computational efficiency of these simulations, this paper proposes a framework with a series of optimizations, including an I/O optimization, several implementation improvements, and two parallelizing schemes. To demonstrate the feasibility and performance of the proposed framework, a Zebiak-Cane (ZC) case model is studied and presented in this paper. Experimental results show that the proposed framework can greatly improve the computational efficiency of climate simulations.