Designing optimal parallel volume rendering algorithms

TL;DR: A methodology to control the complexity in designing parallel algorithms is introduced, and the result is parallel algorithms with all of the features of sequential ones that deliver the promise of parallelism.

Abstract: Volume rendering is a method for visualizing volumes of sampled data such as CT, MRI, and finite element simulations. Visualization of medical and simulation data improves understanding and interpretation, but volume rendering is expensive and each frame takes from minutes to hours to calculate. Parallel computers provide the potential for interactive volume rendering, but parallel algorithms have not matched sequential algorithm's features, nor have they provided the speedup possible. I introduce a methodology to control the complexity in designing parallel algorithms, and apply this methodology to volume rendering. The result is parallel algorithms with all of the features of sequential ones that deliver the promise of parallelism. My algorithms are sufficiently general to run on single instruction multiple data (SIMD) computers and multiple instruction multiple data (MIMD) computers. Through complexity analysis and performance measurements I show that volume rendering is ideally parallelizeable with linear speedup and low memory overhead.