A Distributed GPU-Based Framework for Real-Time 3D Volume Rendering of Large Astronomical Data Cubes

Abstract

We present a framework to volume-render three-dimensional data cubes interactively using distributed ray-casting and volume-bricking over a cluster of workstations powered by one or more graphics processing units (GPUs) and a multi-core central processing unit (CPU). The main design target for this framework is to provide an in-core visualization solution able to provide three-dimensional interactive views of terabyte-sized data cubes. We tested the presented framework using a computing cluster comprising 64 nodes with a total of 128 GPUs. The framework proved to be scalable to render a 204 GB data cube with an average of 30 frames per second. Our performance analyses also compare the use of NVIDIA Tesla 1060 and 2050 GPU architectures and the effect of increasing the visualization output resolution on the rendering performance. Although our initial focus, as shown in the examples presented in this work, is volume rendering of spectral data cubes from radio astronomy, we contend that our approach has applicability to other disciplines where close to real-time volume rendering of terabyte-order three-dimensional data sets is a requirement.

Memory Efficient Load Balancing for Distributed Large-Scale Volume Rendering Using a Two-Layered Group Structure

“…The worst-case memory usage of a single process when using the k-d tree structure is O(v), where v is the number of voxels in the volume. The risk that a high data imbalance occurs limits the use of k-d tree based dynamic load balancing in large-scale applications, where even small data imbalances can result in some processes running out of memory [5].…”

“…In large-scale applications the data sets could consist of multiple terabytes of data, meaning that even small-scale data transfers can be time consuming and result in some processes exceeding their available amount of memory. As such, many large-scale visualization projects have utilized static techniques [5], [12] or have limited load balancing to equalizing the data distribution, rather than explicitly lowering the total render time [10], [13]. Dorier et al [13] developed a technique which in-situ can identify important data of a simulation and reduce less important data based on a time limit constraint.…”

“…However, dynamically redistributing data can result in large memory imbalances between processes. For large data sets some processes might run out of memory, making many dynamic Manuscript load balancing techniques unsuitable for large-scale visualization [4], [5]. Commonly used dynamic load balancing techniques are based on tree structures, e.g., a k-d tree [6].…”

A computational framework for regional seismic simulation of buildings with multiple fidelity models

“…With adequate computational resources, distributed computing can be highly efficient to handle many large scale scientific problems [27][28][29][30][31][32] . For instance, Wijerathne et al [33] used a cluster of workstations to simulate the seismic damage of buildings in Tokyo.…”

Visualizing Astronomical Data with Blender

“…Visualizing and animating output for inspection or study can be run as parallel processes or a graphics processing unit (GPU; Hassan et al 2012). Once keyframes have been inserted, frames generated from the render process become independent.…”