Hybrid-Parallel Simulations and Visualisations of Real Flood and Tsunami Events Using Unstructured Meshes on High-Performance Cluster Systems

Abstract

We present simulations of real flood and tsunami events using a hybrid OpenMP-MPI model on high-performance cluster systems. The two-dimensional shallow water equations were solved by means of the in-house code NUFSAW2D, using an edge-based cell-centred finite volume method with the central-upwind scheme for millions of unstructured cells, thus ensuring spatial accuracy, especially near buildings or hydraulic structures. Each node of a cluster system performed simulations using OpenMP and communicated with other nodes using MPI. We explain strategies on reordering the meshes to support contiguous memory access patterns and to minimise communication cost; to this end, a simple criterion was proposed to decide the strategy used. Despite employing static domain decompositions for such unstructured meshes, the computation loads were distributed dynamically based on the complexity level, to each core and node during runtime to ensure computational efficiency. Our model was tested by simulating two real-life cases: the 2011 flood event in Kulmbach (Germany) and the Japan 2011 tsunami recorded in Hilo Harbour, Hawaii (USA). The numerical results show that our model is robust and accurate when simulating such complex flood phenomena, while the hybrid parallelisation concept proposed proves to be quite efficient. We also provide an outlook for an advanced visualisation method employing the Sliding Window technique with an HDF5 data structure. With such a combination of high-performance computing and interactive visualisation, users have a comprehensive predictive tool to take immediate measures and to support decision makers in developing a well-integrated early warning system.