A GPU Implementation for Two-Dimensional Shallow Water Modeling

In this paper, we present a GPU implementation of a two-dimensional shallow water model. Water simulations are useful for modeling floods, river/reservoir behavior, and dam break scenarios. Our GPU implementation shows vast performance improvements over the original Fortran implementation. By taking advantage of the GPU, researchers and engineers will be able to study water systems more efficiently and in greater detail.

💡 Research Summary

The paper presents a CUDA‑based GPU implementation of a two‑dimensional shallow‑water model originally written in Fortran. The authors adopt Owen Ransom’s 2‑D predictor‑corrector MacCormack scheme, which divides each time step into 16 sub‑steps to respect data dependencies while keeping cell updates independent within a sub‑step. Two GPU versions are developed: a non‑shared‑memory version that launches multiple kernels per time step and stores all data in global memory, and a shared‑memory version that groups threads into 16 × 16 blocks, copies each block’s interior and halo cells into fast on‑chip shared memory, and thus reduces global memory traffic. The non‑shared version suffers from a serial time‑step‑size calculation performed on the CPU, incurring PCI‑Express transfer overhead, whereas the shared version achieves an additional 10–15 % speedup by reusing data locally. Performance is evaluated on an Ubuntu 12.04 system with an Intel i7‑3370K CPU and an NVIDIA GTX 680 GPU using five datasets of varying grid sizes (up to 1,048,576 cells) and simulation lengths. Results show that the GPU code outperforms the original Fortran code by more than an order of magnitude, with larger grids yielding greater relative gains. The shared‑memory optimization further improves runtime, and the authors discuss how increased shared‑memory capacity or larger block sizes could provide even higher speedups. Related work is surveyed, highlighting alternative schemes such as Kurganov‑Petrov flux calculations and early‑exit optimizations for dry cells. The paper also outlines future directions: scaling the solver across multiple GPUs by domain decomposition, handling dry cells to avoid negative water elevations, and exploiting newer heterogeneous CPU‑GPU architectures. Overall, the study demonstrates that explicit shallow‑water simulations are well‑suited to massively parallel GPU hardware, delivering substantial runtime reductions and opening the door to higher‑resolution, more detailed flood and dam‑break analyses.