Large-Scale Computations of Ship Motions with Dynamic Overset Curvilinear Grids

摘要

The advent of faster parallel computers offers both an opportunity and a challenge for the ship hydrodynamics CFD community. On one hand, larger grids allow the modeling of more complex phenomena and resolution of smaller scales. But the recent trend toward massively parallel computers as opposed to faster single processors imposes a significant burden of finding extra concurrency in CFD codes to handle large cases. The purpose of this paper is to present the results of an ongoing effort to improve the serial arid parallel performance of the overset curvilinear grid based, single-phase level set code CFDShip-Iowa version 4, as well as the ability to Jiandle large-scale problems. In addition, demonstration of the improvement in the solutions achieved with high resolution grids is presented for two cases previously computed with grids that can be considered medium refinement for today's standards in ship flows: the forward speed diffraction and the pitch and heave problems in head waves, both for the surface combatant model DTMB 5512. Intensive profiling was added to the code to detect bottlenecks, and several cases ranging from 7 million to 70.6 million grid points were run to subject the code to diverse operational conditions, including cases with 0DOF (static), 2DOF, 6DOF and 6DOF with moving appendages and propeller. These preliminary evaluations revealed several problems in inter- processor information exchange that prevented the code from running in hundreds of processors. Other source of excessive execution time was the overset interpolation process, which currently has poor scalability but that has been considerably optimized for serial use. The overall goal to extend the capability of complex motion cases to over 70 million grid points, and to scale the code to about 500 processors has been achieved for weak scaling, with the exception of I/O that still requires work, due to problems with vendor supplied MPI 10 libraries. Full motions capability for large cases is currently limited by the overset assembly process. A complete set of metrics characterizing the performance of the code are presented.