ParBreZo: A parallel, unstructured grid, Godunov-type, shallow-water code for high-resolution flood inundation modeling at the regional scale

Abstract

Topographic data are increasingly available at high resolutions (<10 m) over large spatial extents to support detailed flood inundation modeling and loss estimation analyses required for flood risk management. This paper describes ParBreZo, the parallel implementation of a two-dimensional, Godunov-type, shallow-water code, to address the computational demand of high-resolution flood modeling at the regional scale (10 2–10 4 km 2). A systematic approach to unstructured grid partitioning (domain decomposition) is presented, and the Single Process Multiple Data (SPMD) paradigm of distributed-memory parallelism is implemented so the code can be executed on computer clusters with distributed memory, shared memory, or some combination of the two (now common with multi-core architectures). In a fully-wetted, load-balanced test problem, the code scales very well with a parallel efficiency of close to 100% on up to 512 processes (maximum tested). A weighted grid partitioning is used to partially address the load balancing challenge posed by partially wetted domains germane to flooding applications, where the flood extent varies over time, while the partitioning remains static. An urban dam-break flood test problem shows that weighted partitions achieve a parallel efficiency exceeding 70% using up to 48 processes. This corresponds to a 97% reduction in execution time so results are obtained in a matter of minutes, which is attractive for routine engineering analyses. A hurricane storm surge test problem shows that a 10 m resolution, 12 h inundation forecast for a 40 km length of coastline can be completed in under 2 h using 512 processors. Hence, if coupled to a hurricane forecast system capable of resolving storm surge, inundation forecasts could be made at 10 m resolution with at least a 10 h lead time.