| The research in this dissertation aims to address the challenges to visualization resulting from large and complex datasets. The thesis is that effective high performance visualization, which is responsive to the challenges of large data, follows from a combination of parallel software architectures and optimizations along with steps to reduce processing load in the visualization pipeline. Broadly speaking, the research follows a bifurcated approach. One approach focuses on algorithms and architectures that leverage parallel computing platforms to increase the capacity of the visualization pipeline. Topics in this approach include a sort-first parallel rendering architecture, remote parallel visualization, and the first-ever study of hybrid parallelization of volume rendering at extreme concurrency. The other approach aims to reduce the amount of work entering the visualization pipeline. We coin the term "query-driven visualization" to refer to the process of limiting the amount of data entering the visualization pipeline to that deemed "scientifically interesting," and use state-of-the-art indexing algorithms for rapid data subsetting. This approach has better performance than the best similar algorithms in visualization, and proves useful in diverse applications like forensic cybersecurity analysis and study of output from a high energy physics plasma laser-wakefield simulation code. |
