Numerical geophysical fluid dynamics models are large parallel applications that need to simulate complex physical phenomena within tight production schedules. In search of improving prediction accuracy, they are continuously exploring new discretization methods and increasing the spatial resolution, which results in an ever increasing demand for more computing power and efficient codes that can leverage non traditional computing architectures. Porting large GFD codes to heterogeneous architectures preserving a single, efficient source code can be a daunting process.
A domain specific language solves the problem by providing a language for expressing GFD codes, that abstracts away optimization details specific to the target computing architectures. Several projects like the GridTools and ESCAPE have developed DSLs for high performance weather applications, with an emphasis on accelerators. We show recent developments and ideas of a new toolchain that allows to develop higher-level frontend DSLs on top of existing efforts. The DSL toolchain aims at providing a concise language for the domain of GFD models, abstracting a large set of optimizers.
21.11.2017
15:15 h