A Programmable Streaming Framework for Extreme-Scale Scientific Visualizations
- Authors
- Publication Date
- Jan 01, 2024
- Source
- eScholarship - University of California
- Keywords
- Language
- English
- License
- Unknown
- External links
Abstract
Emerging computational and acquisition technologies are empowering scientists to conduct simulations and experiments on an unprecedented scale. These advancements can push the frontiers of science and technology with groundbreaking discoveries. However, they also pose significant challenges to traditional scientific visualization workflows. Firstly, the data generated by modern scientific studies using these technologies tends to be extremely large and complex, often resulting in slow processing and rendering times. This demands the development of visualization algorithms that can effectively scale with the size of the data. Secondly, state-of-the-art simulations and experiments produce data at extraordinary rates, complicating the task of generating valuable visualization results for scientists. Therefore, there's a pressing need for more adaptive and intelligent visualization workflows. Lastly, although new computer hardware and architecture can speed up the visualization process, significant performance variations still exist among visualization algorithms due to differing design choices. As a result, optimizing algorithms to better leverage emerging hardware features for enhanced efficiency remains an ongoing necessity.This dissertation addresses the aforementioned challenges by introducing a programmable streaming framework enhanced with implicit neural representation, designed for visualizing extreme-scale scientific data. Specifically, it unfolds three innovative methodologies:Firstly, the framework offers a reactive and declarative programming language for streamlining image generation, layout and interaction creation, and I/O processes, eliminating the need for users to manually control all visualization parameters and procedures. This language enables scientists to define highly adaptive visualization workflows through high-level, rule-based grammars. The system then automatically optimizes the low-level implementation according to these specifications, facilitating the creation of more efficient visualization workflows with simpler coding.Secondly, the framework features a scalable, hardware-accelerated streaming visualization system that allows visualization processes to run concurrently with I/O operations. This system not only achieves state-of-the-art scalability but can also effectively manages complex, multi-resolution data structures. It delivers accurate rendering outcomes, reduces memory usage, and leverages emerging hardware capabilities more efficiently.Finally, the framework integrates implicit neural representation (INR) techniques for data compression and interactive visualization. The use of INRs significantly reduces data size while preserving high-frequency details. Additionally, it enables direct access to spatial locations at any desired resolution, obviating the need for decompression or interpolation.In summary, this dissertation research addresses long-standing challenges inherent in extreme-scale scientific visualization by introducing novel designs and methodologies. The presented framework not only enables more efficient and adaptive visualization workflows but also leverages the latest hardware acceleration and data compression techniques. The implications of these advancements extend beyond mere technical improvements; they pave the way for deeper insights and discoveries across a broad spectrum of scientific studies. This research, therefore, represents a significant leap forward, with the potential to transform the landscape of scientific visualization.