Florian Prill Deutscher Wetterdienst (DWD)
Friday 28th September 2012 - 12:00 to 12:25
INI Seminar Room 1
Novel Optimisation Techniques
Simulation in numerical weather prediction and climate forecasting has a fast-growing demand for memory capacity and processing speed. For the last decade, however, computer technology has shifted towards multi-core chip designs while at the same time on-chip clock rates have increased only moderately. The parallel implementation of the ICON model's nonhydrostatic dynamical core therefore follows a hybrid distributed/shared memory approach, based on the Message Passing Interface (MPI) and the OpenMP API. The ICON code couples the different encapsulated components of the earth system model, e.g. dynamics, soil, radiation, and ocean, with high-level language constructs. Its communication characteristics and programming patterns are designed to meet the main challenges in high performance computing, i.e. load balancing, cache efficiency, and low-latency networking, and take the unstructured triangular C-grid into account, which implies indirect addressing. Besides basic optimization strategies such as loop tiling and grid point reordering, the implementation employs special domain decomposition heuristics, parallel range-searching algorithms with logarithmic complexity, and makes use of asynchronous I/O servers to deal with the potentially prohibitive amount of data generated by earth system models. This facilitates the ICON code to extract an adequate level of performance on a wide range of HPC platforms, targeting large scalar cluster systems with thousands of cores as well as vector computers.