GungHo - a next generation atmospheric dynamical core for weather and climate modelling
To be able to run effectively on the next generation of supercomputers, future atmospheric dynamical cores will need to scale on hundreds of thousands of processors.
The challenge is to achieve the required scalability while retaining the accuracy of the current dynamical core. Designing such a core, therefore, requires a mix of numerical analysis, geophysical fluid dynamics and computational science.
To bring this expertise together, the Met Office, NERC and STFC implemented a five-year program (2011-2016), to research, design and develop a new dynamical core.
Some of the key aspects
- The grid that the model is discretized on. Removal of the singularity in the current latitude-longitude grid is considered essential to achieving good scalability. Although no alternative grid is without its own issues, the cubed-sphere grid has a number of advantages over other choices and is currently the preferred option.
- Highly scalable implicit solvers. There are significant advantages to retaining a two-time-level implicit temporal discretization, but this is only viable if the resulting implicit system, with its global connectivity, can be efficiently solved on hundreds of thousands of processors.
- Inherently conservative advection schemes. Only dry mass is inherently conserved by the current dynamical core, yet there is a growing need to exactly conserve a number of tracer fields, as well as possibly such quantities as energy and angular momentum. This requires replacement of the current pointwise semi-Lagrangian scheme with a flux-form conserving advection scheme, be that a semi-Lagrangian one or an Eulerian one, while preserving the good phase properties of the current scheme.
- The spatial discretization. A mixed finite-element spatial discretization, as distinct from the current finite-difference/finite-volume approach, permits the use of alternative grid structures without some of the disadvantages that those grids incur with a finite-difference discretization.
- A new modelling infrastructure has been designed to permit the efficient implementation of these changes. This is called LFRic after Lewis Fry Richardson (see Related pages).
A further important element is how each of the above interacts with, and depends upon, each other.
- To design and develop a dynamical core that scales well on hundreds of thousands of processors while maintaining at least the accuracy and robustness of its contemporary dynamical core.
- To improve the conservation properties of the dynamical core.