Accelerating Tsunami Computation on a GPU Using OpenMP Offloading

By employing the OpenMP offloading technique, we use GPU to accelerate the numerical wave propagation model TsunAWI. Tsunami computations help us to map and analyse the effects of natural disasters. The numerical model code for the pre-calculation, hindcast, and analysis of tsunami events is designed by mathematical modelling of the corresponding partial differential equations. Implementation and optimisation of the model on HPC systems are especially useful in the context of early warning systems where the fast generation of model estimates is crucial, and acceleration techniques might be especially beneficial. Since the initial code is written in FORTRAN, we didn’t want to rewrite the entire code in CUDA or OpenCL (low-level programming model), so we chose to use OpenMP GPU offloading methodology. For performance study analysis, we used two kinds of meshes; coarser mesh and finer mesh. For both meshes, we see a similar speedup of about 40–50% in GPU compared to CPU (using up to eight threads). Plus, it took less effort to accelerate the TsunAWI code, thus promising that most of the scientific codes written in FORTRAN can easily be accelerated on the exascale machine; most of the exascale machine consists of accelerators.