Session

DescriptionNumerical weather prediction and climate studies rely on efficient algorithms to deliver accurate simulations under tight operational constraints, as increasing horizontal resolutions approach few kilometers in global models and test the performance of legacy codes on massively parallel computing architectures. In this context, time-explicit discretizations are impractical as their time step size is constrained by meteorologically insignificant fast waves. Therefore, implicit or semi-implicit schemes are employed, requiring the solution of large systems of equations that takes up a sizeable portion of computing time. The minisymposium will bring together numerical modellers and high-performance computing experts to explore resiliency and scalability of current strategies for linear solvers in atmospheric models. The session will focus on fault-tolerance, necessary in a context where hundreds of thousands of cores are employed for round-the-clock simulations, and efficiency. Contributions on reduced precision in existing implementations will be surveyed, as well as efforts in supplying solvers with suitable preconditioners that can accelerate convergence. Thus, the talks will give perspectives into modelling approaches that can optimize computational resources and provide fault-robust simulations without sacrificing the accuracy of forecasts produced by existing models.