Icosahedral-grid Models for Exascale Earth System Simulations - ICOMEX
Zusammenfassung der Projektergebnisse
The goal of this project was to make a first step towards preparing the four participating icosahedral Earthsystem models - ICON, MPAS, NICAM and DYNAMICO - for efficient use of future exascale computing platforms. To this end, a selection of key aspects, which are already becoming critical on currently available systems, was addressed, with the intention to develop as generic as possible solutions that will ultimately be beneficial for all participating models. The project was subdivided into seven work packages, most of them focusing either on optimizing the CPU-level computing efficiency or on optimizing several aspects of input/output. In addition, there was an overarching model intercomparison project, addressing both computational and scientific aspects, and a supporting work package dealing with vendor communication. The main achievements and conclusions are as follows: • The model intercomparison study revealed that, from a scientific point of view, the differences among the participating model codes are relatively minor. However, porting complex atmospheric model codes to novel supercomputer architectures including platform-specific optimizations becomes increasingly difficult and time-consuming. Thus, our high expectations in learning from the intercomparison study about general advantages or disadvantages of different coding concepts have not been fully met. • The work packages dealing with improving the CPU-level efficiency of our model codes were a great success. Adapting the memory layout of the model arrays to the optimal setup of a given platform has been found to speed up the execution time by more than a factor of two in some cases. The domainspecific language (DSL) approach pursued here allows to do so without code duplication, which is a mandatory requirement for practicable use in scientific model development. The work on designing efficient implicit solvers of the so-called Helmholtz problem that do not require global communication successfully demonstrated that implicit solvers can be competitive with explicit solvers on massively parallel machines. • The work packages considering input/output (I/O) optimization have also led to significant achievements. As a backbone for improved parallel internal postprocessing, an efficient, scalable, conservative and second-order accurate remapping algorithm has been developed - so far, no comparable algorithm providing all these properties at once is available. The hardware-level work on I/O optimization and parallelization has detected major performance flaws in the implementation of NetCDF and MPI-IO, which have been reported to the developers, and developed an optimized lossless compression algorithm whose compression rate is superior to all comparable existing software packages. Based on our experience, we envision to decouple file format and I/O further. This approach would allow exploiting system-specific characteristics better.
Projektbezogene Publikationen (Auswahl)
- ICON DSL: A Domain Specific Language for Climate Modeling, International Workshop on Domain-Specific Languages and High-Level Frameworks for High Performance Computing (WOLFHPC)
Torres, R., Lindarkis, L., Kunkel. J., and Ludwig, T.
- Reducing the HPC-Datastorage Footprint with MAFISC - Multidimensional Adaptive Filtering Improved Scientifc data Compression, in: Computer Science - Research and Development. Springer, Hamburg. Berlin, Heidelberg. 2012
Hubbe, N., and J. Kunkel
(Siehe online unter https://dx.doi.org/10.1007/s00450-012-0222-4) - Evaluating Lossy Compression on Climate Data. In: Supercomputing, Lecture Notes in Computer Science (7905), pp. 343-356, (Editors; Julian Martin Kunkel, Thomas Ludwig, Hans Wenner Meuer), Springer (Berlin, Heidelberg). ISC 2013. Leipzig, Germany, ISBN: 978-3-642-38749-4
Hubbe, N., Wegener, A., Kunkel, J., Ling, Y. and, Ludwig, T.
- Source-to-Source Translation for Climate Models, Leipzig, Germany, International Supercomputing Conference 2013, 2013-06-17
Torres, R., Lindarkis, L., and Kunkel, J.
