Dipoma Thesis - Parallel GPU-based Vectorfield Visualization on Curved Surfaces


The diploma thesis is available for download - please note that it is written in German.


Recently, an approach was developed which allows the visualization of three dimensional vector fields on standard consumer hardware. However, the computational cost for this approach is very high. The resulting frame rates are usually below interactive rates. Therefore, this diploma thesis parallelizes this approach and optimizes it for use on a graphics cluster. Two different techniques were implemented which distribute the workload evenly on all available cluster nodes in order to achieve the highest possible rendering speed. Additionally, an object-space partitioning of the vector field is implemented, which allows to load huge flow data sets.

Quick Overview

viewport divided in stripes

To speed up the rendering of a scene, the viewport is split up in stripes. Each of these stripes is rendered separately by one of the cluster nodes. Two different load balancing algorithms were developed, that manage the size of these stripes according to the displayed scene. The example image to the left shows the teapot scene, which is rendered using four cluster nodes (stripes inserted to visualize the concept - these stripes are invisible in the actual program).

vector field divided into bricks

An additional goal of the thesis was to allow the rendering of huge vector fields. To achieve this, object-space partitioning was implemented which distributes parts of the vector field over the available cluster nodes. This concept of "bricking" is also used in volume rendering applications. Using all eight cluster nodes allows loading of vector fields with an approx. size of 1,5 GB. The basic idea is shown in the image to the left, where four cluster nodes are used (The red cube symbolizes the whole vector field which is divided into four bricks, indicated by the white lines.)

Last modified: Jan. 25th, 2007
by Sven Bachthaler