Acceleration of a Meteorological Limited Area Model with Dataflow Engines

Climate and weather modeling is a significant consumer of High Performance Computing due to the hard deadlines inherent in predicting weather. Given the large data volumes and runtimes involved, climate and weather modeling is ideally suited for dataflow computation. In this paper, we demonstrate a dataflow implementation of the dynamic core of a meteorological limited area model. To achieve maximum performance we transform the computation by reordering operations and encoding the data. We present results for a domain of 13,600 x 3,333 × 30 km with 620 thousand grid points, and show speedups of up to 74x comparing an x86 CPU node to a dataflow node.