Inland ships continuously operate in restricted waters, where the depth and width are regularly less than twice the ship's draft and four times ship breadth, respectively. In restricted water, the flow around the hull changes compared to that in unrestricted water due to presence of the fairway bottom and sides, that lead to increased return flow, stronger squat effects and changes in the wave pattern produced by the ship. If these changes to the flow are significant, it is worthwhile to optimize the hull form for shallow or confined water rather than for unrestricted water. This paper specifically focuses on the effects of water depth on inland ship stern optimization. It presents the optimization of propulsion power for various water depths using a parametric inland ship stern shape, CFD and surrogate modeling. The change of parameter influence in different water depths is analyzed and explained by means of flow visualization. Using Pareto fronts, a trade-off is shown: propulsion power in shallow water can be decreased at the cost of increased propulsion power in deep water and vice versa.@en

Preconditioners based on incomplete factorization are very popular for a fast convergence of the PCG-algorithm. However, these preconditioners are hard to parallelize since most operations are inherently sequential. In this paper we present the RRB-solver, which is a PCG-type solver using an incomplete Cholesky factorization based on the Repeated Red-Black (RRB) method. The RRB-solver scales nearly as well as Multigrid, and in this paper we show that this method can be parallelized very efficiently on modern computing architectures including GPUs. For an efficient parallel implementation a clever storage scheme turns out to be the key. The storage scheme is called r1=r2=b1=b2 and it ensures that memory transfers are coalesced throughout the algorithm, yielding near-optimal performance of the RRB-solver. The r1=r2=b1=b2-storage scheme in combination with a CUDA implementation on the GPU gives speedup factors of more than 30 compared to a sequential implementation on one CPU core for 5-/9-point stencils problems. A comparison with algebraic Multigrid further shows that the RRB-solver can be implemented very efficiently on the GPU.@en

Preconditioners based on incomplete factorization are very popular for a fast convergence of the PCG-algorithm. However, these preconditioners are hard to parallelize since most operations are inherently sequential. In this paper we present the RRB-solver, which is a PCG-type solver using an incomplete Cholesky factorization based on the Repeated Red-Black (RRB) method. The RRB-solver scales nearly as well as Multigrid, and in this paper we show that this method can be parallelized very efficiently on modern computing architectures including GPUs. For an efficient parallel implementation a clever storage scheme turns out to be the key. The storage scheme is called r1=r2=b1=b2 and it ensures that memory transfers are coalesced throughout the algorithm, yielding near-optimal performance of the RRB-solver. The r1=r2=b1=b2-storage scheme in combination with a CUDA implementation on the GPU gives speedup factors of more than 30 compared to a sequential implementation on one CPU core for 5-/9-point stencils problems. A comparison with algebraic Multigrid further shows that the RRB-solver can be implemented very efficiently on the GPU.@en