Distributed GPU Volume Rendering of ASKAP Spectral Data Cubes

Hassan, Amr

The Australian SKA Pathfinder (ASKAP) will be producing 2 terabyte spectral-line cubes by 2013. Global views of spectral data cubes are vital for the detection of instrumentation errors, the identification of data artefacts and noise characteristics, and the discovery of strange phenomena, unexpected relations, or unknown patterns. We present the first framework that can render ASKAP-sized cubes at interactive frame rates. The framework provides the user with a real-time interactive volume rendering by combining shared and distributed memory architectures, distributed CPUs and graphics processing units (GPUs), using the ray-casting algorithm.

The framework was tested on a cluster with up to 64 GPUs, and was capable of rendering 200 GB spectral data cube at 5 fps. We expect our framework will maintain this performance for terabyte-sized data cubes (with additional hardware). Moreover, the integrated remote visualization capability can provide such a facility to geographically distributed astronomers with average computers and internet connections. The framework can also use a tiled-display system to provide a high resolution rendering output. We demonstrate the ability to integrate such a facility with existing automated data analysis tools by combining the output of Duchamp (a candidate source finder for proposed ASKAP surveys) into the visualization output with the ability to inspect the source finding output in place.

Return to oral presentation list