Query-driven Multiscale Data Postprocessing in Computational Fluid Dynamics

Massively parallel computational uid dynamics codes that have to stream solution data to a visualisation or postprocessing component in each time step often are IO-bounded. This is especially cumbersome if the succeeding components require the simulation data only in a coarse resolution or only in specific subregions. We suggest to replace the streaming data approach found in many applications with a query-driven communication paradigm where the postprocessing components explicitly inform the uid solver which data they need in which resolution in which subregions. Two case studies reveal that such a data exchange paradigm reduces the memory footprint of the exchanged data as well as the latency of the data delivery, and that the approach scales. In particular geometric multigrid solvers based upon a non-overlapping domain decomposition can answer such queries efficiently