Portable multi-GPU urban flood modelling via MPI-OpenACC: evaluating acceleration, scaling, and workload-per-GPU

We present a multi-GPU implementation of the physics-based 2D urban flood model (H12) utilising MPI-OpenACC, demonstrating significant reductions in runtime for high-resolution, city-scale, and large-basin simulations. Applied to grids ranging from 1 to 20 m across two basins (downtown Portland, Oregon, USA; downstream Han River, South Korea), the multi-GPU configuration achieved a speedup of approximately 710× over a serial baseline and approximately 30× over a 144 core CPU-hybrid run. This advancement renders metre-scale, city-wide scenarios and extensive domains computationally feasible. Our findings provide practical guidance for determining optimal GPU counts based on workload per device; while exact thresholds vary with problem size (spatial coverage and resolution) and system interconnect, the underlying principle aids in aligning domains with available accelerators. The directive-based MPI-OpenACC design maintains a unified code base across CPU and GPU execution, thereby supporting portability and operational use. • We present a portable CPU–GPU urban flood model using MPI-OpenACC. • Multi-GPU cuts runtime by up to 710× (serial) and 30× vs multi-CPU. • The model enables tractable 1-20 m city-scale and large-domain simulations. • We provide work-per-device guidance to choose GPU counts across problem sizes.