| Massively parallel supercomputers enable scientists to simulate complex phenomena in unprecedented detail. When scientists attempt to analyze and understand data generated by large-scale simulations, the sheer size of the data is a major challenge. Many advances have been made for large-scale data visualization that address this challenge. However, they do not always meet the performance requirements of interactive, high-resolution/high-precision visualization. Visualizing these terascale data sets, soon to be petascale, needs the storage space and processing power of the same supercomputers used for the simulations.;In this dissertation, we design, implement, and evaluate parallel visualization pipelines using high-performance parallel supercomputers. We carefully study the interplay between data representation, storage organization, hardware architecture, access pattern of a variety visualization operations, user interaction, and application requirement. A set of effective methods is presented, eliminating the key bottlenecks that arise in interprocessor communication, data partition and distribution, data transferring, and data preprocessing. The feasibility and scalability of our approaches have been demonstrated with a performance study using thousands of processors on different supercomputer systems. Our approaches have been applied to effectively and efficiently visualize several scientific simulations, from earthquake modeling to supernova evolution, at the highest resolution to date. This high-performance parallel visualization capability significantly enhances scientists' power to explore and understand large, complex, time-varying volume data. |
