Efficiency improvement of viscous ship flow computations through use of the Graphics Processing Unit

Efficiency improvement of viscous ship flow computations through use of the Graphics Processing Unit: A performance analysis on different hardware

de Bruycker, Deborah (TU Delft Aerospace Engineering)

van Zuijlen, A.H. (mentor)

Delft University of Technology

Aerospace Engineering

2017-09-28

Maritime hydrodynamics involves strong inertia-driven flows, including free-surface waves, and with Reynolds numbers as high as 109. Numerical modelling of these flows is therefore a true challenge. Especially for the optimization of the design of the aft part of the ship, where viscous effects cannot be discarded. At MARIN, the viscous flow solver PARNASSOS has been developed for this purpose. Over the years, this solver has been optimized with respect to robustness, accuracy and efficiency, which has resulted in a tool capable of doing fast viscous flow computations. However, if the ship's generated wave pattern has to be taken into account, a complete hull form evaluation can still take hours, or even days. For automatic optimization purposes, where hundreds or thousands of calculations are needed, it is thus desired to further accelerate the computing time of this solver. Most of the CPU time for a computation with PARNASSOS is spent on solving the large sparse linear systems of the form Ax=b. These are currently solved using an iterative solver such as GMRES, in combination with a preconditioner to improve convergence rates. Due to the increase in computational capability of modern computers, the performance of the linear system solver could be further improved through use of high performance computing. Especially the Graphical Processing Unit (GPU) seems to have great potential for speeding up such scientific computations. The aim of the work is therefore to investigate whether it is possible to achieve reasonable speedup of PARNASSOS' linear system solver by making use of GPU computing. This investigation was done using both model- and full-scale test cases, both for increasing size of the systems to solve. Performance of different iterative solvers has been analysed on different GPU cards and has been compared against the performance, currently used on the CPU.

GPU computing
Preconditioning
CFD
Krylov solvers

http://resolver.tudelft.nl/uuid:87ad4bb3-61b5-4a55-a956-1db361f133c1

Student theses

master thesis

© 2017 Deborah de Bruycker