|
This work presents a system for the hardware accelerated parallel
rendering of large volume datasets on PC Clusters. The
parallelization possibilities on the visualization pipeline are
explained, and how some of this possibilities have been exploited.
The evolution of commodity graphics hardware is also discussed, and
serves as motivation for working on the parallelization of the last
step of the visualization pipeline, namely the rendering of images.
A comparison between ''high end'' proprietary
visualization systems and commodity hardware is provided and serves
as motivation for working on the parallelization of the rendering
using clusters of PCs equipped with commodity graphics hardware. The
parallelization of the volume rendering algorithm is presented, and
special attention is paid to the parallelization of the compositing
stage: two well known algorithms for parallel compositing, direct
send and binary swap, are discussed and compared and a justification
for choosing one over the other is provided. Assembler
implementations of the compositing operation are given, one for the
Intel's x86 architecture, and one for AMD's AMD64. The
performance characteristics of this implementations are discussed.
Remote access of generic visualization facilities is presented,
including a detailed description of a generic method for remotely
accessing visualization applications.
|