Distributed Neural Representation for Reactive in situ Visualization

TL;DR

DVNR introduces a distributed volumetric INR that trains on partitioned HPC data via model parallelism to enable in situ, reactive visualization without interprocess data exchanges. By combining ghost-cell boundary handling, boundary losses, and weight caching, DVNR achieves competitive compression quality and scalability while integrating with DIVA and Ascent for real-world workflows. The approach is validated across multiple HPC datasets and in situ simulations, demonstrating memory-efficient temporal caching and interactive visualization capabilities. While effective, the method highlights limits such as automatic INR sizing and the need for more DVNR-friendly visualization operators, pointing to fruitful directions for future work.

Abstract

Implicit neural representations (INRs) have emerged as a powerful tool for compressing large-scale volume data. This opens up new possibilities for in situ visualization. However, the efficient application of INRs to distributed data remains an underexplored area. In this work, we develop a distributed volumetric neural representation and optimize it for in situ visualization. Our technique eliminates data exchanges between processes, achieving state-of-the-art compression speed, quality and ratios. Our technique also enables the implementation of an efficient strategy for caching large-scale simulation data in high temporal frequencies, further facilitating the use of reactive in situ visualization in a wider range of scientific problems. We integrate this system with the Ascent infrastructure and evaluate its performance and usability using real-world simulations.

Figures (27)

• Figure 1: The design of distributed volumetric neural representation (DVNR).
• Figure 2: A) Illustration of ghost cells used in DVNR training (assuming cell-centered volume discretizations diskin2011comparison in the illumination). B) In DVNR training, boundary connectivity is enhanced by utilizing training samples from two distinct distributions: uniform and boundary-centered half-Gaussian. More specifically, $L1$ losses are computed for samples from each distribution, which are then combined through a weighted average.
• Figure 3: This figure illustrates a comparison between boundary slices of two adjacent partitions using different weighting factors (referred to as $\lambda$). It features pseudo-color plots that highlight the differences between the slices. Both neural representations were trained for 10k steps using the S3D heat release field.
• Figure 4: This figure presents volume visualizations of multi-resolution hash encoding weights arranged into 3D grids. A) The visualization at a dense resolution clearly reveals the shape of the underlying object within the volume data. B) At a hashed resolution, the visualization shows some noise, yet the fundamental structure of the object is still recognizable. C) For reference, the volume visualization of the actual data is displayed. D) The diagram outlines our model compression strategy, applying 3D SZ3 compression liang2022sz3 to dense resolutions and 1D ZFP compression to hashed resolutions. In this example, the Mechanical Hand dataset is compressed via a single INR that employs a two-level multi-resolution hash encoding, yielding a 40.7dB PSNR reconstruction quality after compression.
• Figure 5: The overview of our DIVA-Ascent integration.