Performance quantification of visualization systems

Abstract

Visualization is an important part of data analysis, complementing automatic data processing to provide insight in the data and understand the underlying structure or patterns. A visualization system describes a visualization algorithm running on a specific compute architecture or device. Runtime performance is crucial for visualization systems, especially in the context of ever-growing data sizes and complexity. One reason for this is the importance of interactivity, another is to provide the opportunity for a comprehensive investigation of generated data in a limited time frame. Providing the possibility of changing the perspective beyond the original focus has been shown to be particularly helpful for explorative data analysis. Performance optimization is also key to save costs during visualization on supercomputers due to the high demand for their compute time. Being able to predict runtime enables a better resource planning and optimized scheduling on such devices. The central research questions addressed in this thesis are threefold and build on each other: How can we quantify runtime performance of visualization systems? How to use this information to develop models for prediction, and ultimately: How to integrate both aspects in the application context? The goal is to gain a comprehensive understanding of the runtime performance of visualization systems and optimize them to save costs and improve the user experience. Despite many works in this direction, there are still open questions and challenges on how to reach this goal. One of these challenges is the diversity of compute architectures used for visualization, including devices from mobile devices to supercomputers. Most visualization algorithms profit from running in parallel. However, this poses another challenge in performance quantification due to the usage of multiple heterogeneous parallel hardware hierarchies. Typically, visualization algorithms deal with large data, sparse regions, and interactivity requirements. Further, they can be fundamentally different in their rendering approaches. All these aspects make a reliable performance prediction difficult. This thesis addresses those challenges and presents research on performance evaluation, modeling, and prediction of visualization systems, and how to translate these concepts to improve performance-critical applications. Assessing runtime performance plays a key role in understanding and improving it. A new framework for the extensive and systematic performance evaluation of interactive visualizations is introduced, to help gain a deeper understanding of runtime behavior and rendering parameter dependencies. Based on the current practice of runtime performance evaluation in literature, a database of performance measurements is created. A list of best practices on how to improve performance evaluation is compiled based on a statistical analysis of the data. Additionally, a frontend has been developed to visually compare the rendering performance data from multiple perspectives. With a fundamental understanding of an application's runtime behavior, performance can be modeled, and the model used for prediction. New techniques for different hardware systems are introduced that are typically used for the visualization of large data sets: desktop computers featuring dedicated graphics hardware and high-performance distributed memory systems. For the former, a method to predict performance on-line is used to dynamically tune volume rendering during runtime to guarantee interactivity. For image database generation on distributed memory systems, a hybrid approach for dynamic load balancing during in situ visualization is introduced. This work also explores how human perceptual properties can be used to improve the performance of visualization applications. Two novel techniques are introduced that adapt rendering quality to the human visual system by tracking the users gaze and changing the visualization accordingly. In this thesis, a special focus is set on volume rendering. Performance optimization makes it possible to use volume rendering to visualize data outside the typical use cases. Two visualization systems are presented that use volume rendering at their core: one for the interactive exploration of large dynamic graphs and one for the space-time visualization of gaze and stimulus data. Overall, this thesis advances the state of the art by introducing new ways to assess, model, and predict runtime performance of visualization systems that can be used to improve usability and realize cost savings. This is demonstrated through several applications.