Abstract

We address the problem of distributing rendering computations for real-time display of very complex three dimensional (3D) scenes using a graphics cluster. The rendering of 3D scenes is increasingly being carried out using at least two different programs on the graphics processing unit (GPU): a vertex shader program for vertex (geometry) processing, and a fragment shader program for pixel (colour) processing. With fragment shader programs becoming more and more time consuming, distributing load solely based on geometry -- as is done in most contemporary systems -- can cause significant load imbalance and redundant work.

In this thesis we propose a number of parallel rendering algorithms which divide the traditional cluster rendering pipeline into two different phases: one which primarily concerns itself over vertex operations to generate depth information, and a second which primarily concerns itself over fragment operations. By performing communication between these two phases, each node can perform fewer fragment operations with little overhead over traditional cluster rendering algorithms. We also propose a number of load-balancing algorithms which utilize the information gained earlier in the pipeline to improve the management of GPU resources. The techniques are implemented on a graphics cluster and experimental results demonstrate significant improvements in rendering performance.