RmdnCache: Dual-Space Prefetching Neural Network for Large-Scale Volume Visualization.

Area of Science:

• Computer Science
• Data Visualization
• Scientific Computing

Background:

• Large-scale 3D scientific datasets are crucial for pattern discovery.
• Interactive volume visualization faces challenges due to high input latency from I/O bottlenecks and memory limitations.
• Existing systems struggle with seamless user experience for complex datasets.

Purpose of the Study:

• To introduce RmdnCache, a deep learning-based prefetching method.
• To optimize data flow across the memory hierarchy for large-scale volume visualization.
• To reduce input latency and improve interactive visualization performance.

Main Methods:

• Developed a deep learning architecture combining Recurrent Neural Networks (RNN) and Mixture Density Networks (MDN).
• Implemented a prefetching strategy that predicts the next view's content and its likelihood distribution.
• Optimized data flow to fast memory during the rendering of the current view.

Main Results:

• RmdnCache accurately prefetches necessary data into fast memory.
• Significantly reduces overall input latency for large-scale volume visualization.
• Outperforms existing state-of-the-art prefetching algorithms on real-world datasets.

Conclusions:

• RmdnCache effectively addresses latency issues in interactive volume visualization.
• The proposed deep learning approach enhances the user experience for large scientific datasets.
• This method offers a promising solution for efficient processing of complex volumetric data.