Parallel coupling framework and advanced time integration methods for detailed cloud processes in atmospheric models
Final Report Abstract
Clouds and precipitation are of significant importance to the atmosphere. Thus, they influence weather as well as climate. However, cloud processes still represent one of the major uncertainties in current atmospheric forecast models. A class of highly detailed schemes for the description of cloud microphysical process are spectral bin microphysical models. However, their application in atmospheric models is mostly not feasible, primarily due to the high computational demands in comparison to the well-established simplified approaches. The objective of the project has been to investigate advanced numerical schemes and innovative methods of high performance computing as a potential for performance improvements. We developed a new method of coupling spectral cloud models to an atmospheric model. After separating data structures and parallelization of both models, we introduced dynamic load balancing for the complex computations of the spectral cloud model. The costs of these computations are strongly varying and have been a source of load imbalances in previous models. The advantage of this separation is, that no changes in the data structures of the atmospheric model are required to load balance the microphysical computations. We implemented the framework FD4 (Four-Dimensional Distributed Dynamic Data structures) in the course of the project to provide an abstraction for dynamic data management, load balancing, and model coupling. Experiments with the model system COSMO–SPECS+FD4 have shown that dynamic load balancing improves the performance by a factor of two. Furthermore, our approach has demonstrated scalability to 64 thousand processes, which allows an efficient usage of present high-performance computing systems. A large model area of 512×512×48 grid cells could be computed with the spectral microphysics in real-time, which shows that the application for comprehensive studies is technically feasible. The numerical description of clouds involves processes with very different characteristics. Splitting methods can be employed to solve different parts of an evolution equation with an appropriate time integration scheme. For advection-reaction equations so-called IMEX (implicit/explicit) splittings have proven efficient. In the context of cloud modeling the reaction part represents microphysics. Pure transport equations with locally varying time step restrictions can efficiently be solved employing explicit multirate methods. We constructed a generic splitting scheme that allows for the construction of IMEX-multirate methods, combining the advantages of both approaches. The practical use of advanced time integration schemes is often limited by the more complex program flow. The splitting scheme developed in the course of this project can be implemented with little overhead. Sophisticated balancing methods ensure that the implementation scales well with the number of processors, which is crucial for parallel computations. A variety of simulations could potentially benefit from the incorporation of our results. Especially atmospheric models expose a wide spectrum of temporal and spatial scales so that their computational cost can be reduced significantly by employing multirate schemes. The application of more detailed cloud descriptions in future weather models can benefit from the numerical methods and dynamic load balancing methods developed in this project. Furthermore, detailed spectral schemes can act as a benchmark to improve the simpler schemes used in current forecast models.
Publications
- Multirate Runge–Kutta schemes for advection equations. JCAM, 226:345–357, 2009
Martin Schlegel, Oswald Knoth, Martin Arnold, and Ralf Wolke
- A framework for detailed multiphase cloud modeling on HPC systems. In Parallel Computing: From Multicores and GPU’s to Petascale, volume 19, pages 281–288. IOS Press, 2010
Matthias Lieber, Ralf Wolke, Verena Grützun, Matthias S. Müller, and Wolfgang E. Nagel
- FD4: A Framework for Highly Scalable Load Balancing and Coupling of Multiphase Models. AIP Conference Proceedings, 1281(1):1639–1642, 2010
Matthias Lieber, Verena Grützun, Ralf Wolke, Matthias S. Müller, and Wolfgang E. Nagel
- Multirate Implicit- Explicit Time Integration Schemes in Atmospheric Modeling. AIP Conference Proceedings, 1281(1):1831–1834, 2010
Martin Schlegel, Oswald Knoth, Ralf Wolke, and Martin Arnold
