Sort-middle stream-based real-time rendering on commodity clusters /by Jorge Luis Williams.

Today, expensive high-end graphics machines are losing their competitive edge to clusters of inexpensive PCs with commodity graphics accelerator cards. A popular approach for achieving cluster-based rendering is to utilize a stream-based rendering approach. A stream-based system achieves parallel rendering on a PC-cluster by distributing, processing, tracking, and synchronizing streams of API commands (usually OpenGL). There are three basic algorithms for parallelizing the graphics pipeline: sort-first, sort-middle, and sort-last. Both sort-first and sort-last renders have been developed using stream based rendering on clusters. Unfortunately, these two rendering techniques are either difficult to load-balance or difficult to scale as the resolution increases. The sort-middle approach, that has been successful on high-end graphics machines, has been discounted as difficult to attain on cluster systems because of the tight coupling of geometry and rasterization processes on commodity graphics hardware. Because of this, no one has implemented a sort-middle cluster rendering system of any kind.;The overall objective of this research is to design and implement a sort-middle stream-based algorithm for cluster rendering. We wish to examine and quantify the advantages and disadvantages of utilizing such an algorithm in comparison to other rendering algorithms. We utilize standard programmable graphics hardware and modify standard APIs to achieve a separation in the rendering pipeline that enables the implementation of the sort-middle approach. We incorporate this algorithm in an existing stream-based rendering (SBR) system and examine its performance when compared to the system's sort-first implementation. We also iv make comparisons against a sort-last non-stream based system since sort-last performs dramatically better in this case.;Not only do our results indicate that sort-middle is feasible to implement on a cluster, but also that in some cases it has advantages in terms of load-balance and reduced network traffic. This is especially true in cases where dynamic scenes must be rendered in immediate mode or where geometric entities in a scene are unorganized spatially. Additionally, we found that sort-middle is capable of efficiently utilizing nodes that do not support graphics processing units (GPUs) in GPU-limited clusters.Thesis (Ph. D., Computer Science)--University of Idaho, December 2008