Strategies for efficient parallel visualization

Abstract

Visualization is a crucial tool for analyzing data and gaining a deeper understanding of underlying features. In particular, interactive exploration has shown to be indispensable, as it can provide new insights beyond the original focus of analysis. However, efficient interaction requires almost immediate feedback to user input, and achieving this poses a big challenge for the visualization of data that is ever-growing in size and complexity. This motivates the increasing effort in recent years towards high-performance visualization using powerful parallel hardware architectures. The analysis and rendering of large volumetric grids and time-dependent data is particularly challenging. Despite many years of active research, significant improvements are still required to enable the efficient explorative analysis for many use cases and scenarios. In addition, while many diverse kinds of approaches have been introduced to tackle different angles of the issue, no consistent scheme exists to classify previous efforts and to guide further development. This thesis presents research that enables or improves the interactive analysis in various areas of scientific visualization. To begin with, new techniques for the interactive analysis of time-dependent field and particle data are introduced, focusing both on the expressiveness of the visualization and on a structure allowing for efficient parallel computing. Volume rendering is a core technique in scientific visualization, that induces significant costs. In this work, approaches are presented that decrease this cost by means of a new acceleration data structure, and handle it dynamically by adapting the progressive visualization process on-the-fly based on the estimation of spatio-temporal errors. In addition, view-dependent representations are presented that both reduce the size and render cost of volume data with only minor quality impact for a range of camera configurations. Remote and in-situ rendering approaches are discussed for enabling the interactive volume visualization without having to move the actual volume data. In detail, an approach for the integrated adaptive sampling and compression is introduced, as well as a technique allowing for user prioritization of critical results. Computations are further dynamically redistributed to reduce load imbalance. In detail, this encompasses the tackling of divergence issues on GPUs, the adaptation of volume data assigned to each node for rendering in distributed GPU clusters, and the detailed consideration of the different performance characteristics of the components in a heterogeneous system. From these research projects, a variety of generic strategies towards high-performance visualization is extracted, ranging from the parallelization of the program structure and algorithmic optimization, to the efficient execution on parallel hardware architectures. The introduced strategy tree further provides a consistent and comprehensive hierarchical classification of these strategies. It can provide guidance during development to identify and exploit potentials for improving the performance of visualization applications, and it can be used as expressive taxonomy for research on high-performance visualization and computer graphics.