Seminarium LFRic: tackling the exascale challenge for weather and climate modelling 100 years after LF Richardson
Seminarium poświęcone projektowi LFRic mającemu na celu implementację numerycznego modelu pogody następnej generacji, który w perspektywie kilku lat zastąpi obecny model UM. Prelegentami będą Nigel Wood (Met Office, Head of Dynamics Research) oraz Chrisopher Maynard (Met Office).
Abstrakt seminarium:
The joint Met Office, NERC, and STFC GungHo project (2011-2016) proposed a new dynamical core for the Unified Model that targets efficient use of next generation High Performance Computers (HPCs). That new dynamical core is based on a finite-element discretization applied to the cubed-sphere. However, the current Unified Model software infrastructure is unsuitable for such a change to the model and is unlikely to meet the challenges of the next generation of supercomputers. Therefore the GungHo project additionally proposed a new modelling infrastructure. The LFRic project (named after L.F. Richardson) is implementing that design as the replacement for the Unified Model. As well as having to support the finite element data structures required by GungHo, LFRic must be flexible to changes in both HPC architecture and model formulation (and hence is based on unstructured meshes), whilst delivering good computing performance on future exascale supercomputers.
To achieve these goals, the scientific code has been separated from the underlying computational infrastructure, so that changes can be made to either without impacting the other. Code auto-generation techniques are used to optimize the model for different system architectures. This split of responsibility is termed the “Separation of Concerns”. Scaling experiments on the new Met Office supercomputer using auto-generated code confirm the viability of the approach.
This talk will give an overview of the design of both GungHo and LFRic, and the rationale for those designs. The planned time line, and the challenges that that presents us with, will also be discussed.