Master thesis : Discrete Element Simulation of Ice Particle Interaction: Migration to GPU Computing and Subsequent Validation

Abstract

A numerical simulation code using Discrete Element Method has been developed by HSVA, which can generate brash ice and ice ridges as well as analyse ship navigation through ice channels. The current version is simulating the problem in model scale for ease of validation. This thesis aims to enhance the software's capabilities, reduce the computational time, and to enhance performance and capabilities by modifying internal source code. To improve the performance GPU programming has been introduced. GPU programming extension CUDA, developed by NVIDIA, has led to numerous advances in computing over the last few years. The CUDA API makes it relatively easy for users to access the graphics card hardware, which allows users to perform parallel computations with thousands of CUDA cores. The following report investigates the methodology and advantages of using the CUDA API for DEM computations. In order to achieve this, existing CPU code had to be rewritten for the GPU. Both implementations show significant improvements with regard to iteration time, and performance depending on of GPU architecture. Additionally, it has been demonstrated that the GPU can be sped up by simply varying certain parameters, which boosts the code's performance overall. Another investigation dives into the overhead associated with programming memory intensive scripts to the GPU and shows what effect this has on the total calculation times for the application. Further a more complex IceStructure interaction algorithm can improve the quality of results. In this case a different number of simulations is done varying the element number to find out the dependency of the elements to the computational time. Eventually, several tests were conducted for different types of brash ice and Ice ridge channels to see how the software would react.