High-performance computing for long-range underwater acoustics

Abstract

The Comprehensive Nuclear-Test-Ban Treaty Organization (CTBTO) has been recording underwater signals using seven hydrophone stations which cover the entire oceans and provides the data publicly available. One of the lessons learned during the data analysis, such as in the ARA San Juan submarine tragedy, is that underwater sound propagation paths become complex in a long-range propagation, and as such, computer modelling, particularly three-dimensional (3D), can play an important role in understanding sound source location. The biggest stumbling block toward long-range 3D modeling is obviously its huge computational resource demand. In this presentation, we will discuss the GPU implementation of the 3D Split Step Fourier Parabolic Equation (SSFPE) solver, and the domain decomposition parallelization of our newly developed 2D + 1D Finite-Difference Time-Domain (FDTD) solver. The obtained computational performances are: (1) 3D SSFPE on GPU performs approximately 20 times faster than the original Matlab code and (2) the parallel speedup of the FDTD solver with domain decomposition up to 4 nodes is 98%. Although the performances on a larger system are still of interest, those results already enable us to tackle realistic problems. Results highlight the importance of using high performance computing for realistic global scale underwater acoustic simulations.