Large-scale Ultrasound Simulations using Accelerated Clusters
Efficient utilization of accelerated HPC clusters is particularly sensitive to communication efficiency of deployed algorithms. In this thesis, we reexamine pseudo-spectral solvers for wave-like problems in medical ultrasonics to allow their deployment on these machines. The domain decomposition is shown to be a preferable approach to improving data locality of these solvers as a range of suitable alternative discretizations exhibited considerably worse numerical properties.The local Fourier basis domain decomposition is then used to construct a novel solver based on the state of the art model for ultrasound in medicine -- k-Wave. We show that this approach is up to 7.5x faster and achieves almost perfect weak-scaling up to 512 GPU accelerated nodes, while being able to take full advantage of advanced GPU interconnects such as NVLink in NVIDIA DGX-2 multi-GPU nodes. The method offers flexible accuracy--efficiency trade off, which allows to nearly match accuracy of the global k-Space method or maximize performance at sufficient accuracy by subdomain overlap scaling.
pseudo-spectral methods, supercomputing, domain decomposition methods, local Fourier basis, medical ultrasound simulation, non-linear ultrasound, HIFU, MPI+X, CUDA, GPU, accelerated computing