HR: 1340h
AN: IN43B-1182 [Abstracts]
TI: Enabling GPU Acceleration of HIRDLS Scientific Data Processing
AU: Craft, J V
EM: jcraft@ucar.edu
AF: University of Colorado
Center for Limb Atmospheric Sounding, 3450 Mitchell Lane
Building FL-0, Boulder, CO 80301, United States
AU: * Ellis, D T
EM: dtellis@ucar.edu
AF: University of Colorado
Center for Limb Atmospheric Sounding, 3450 Mitchell Lane
Building FL-0, Boulder, CO 80301, United States
AU: Cavanaugh, C
EM: cavanaug@ucar.edu
AF: National Center for Atmospheric Research, 3450 Mitchell Lane
Building FL-0, Boulder, CO 80301, United States
AU: Gille, J G
EM: gille@ucar.edu
AF: University of Colorado
Center for Limb Atmospheric Sounding, 3450 Mitchell Lane
Building FL-0, Boulder, CO 80301, United States
AU: Barnett, J J
EM: j.barnett1@physics.ox.ac.uk
AF: Oxford University
Department of Atmospheric Physics, Clarendon Lab
Parks Road, Oxford, OX1 3PU, United Kingdom
AB:
The HIRDLS team involves scientists, software developers and students from Atmospheric Science, Computer
Science and various engineering disciplines, and from numerous agencies, universities, and centers. The focus
of research is determining vertical profiles of the Earth's atmospheric chemical composition from infrared
samplings.
The processing techniques used to calculate the scientific data rely heavily on mathematical algorithms. While
the data processing involves large sets of data, there is much repetition of smaller well defined kernels. Currently
all processing has been done on large, expensive, CPU clusters. Programming these machines effectively to
take advantage of obvious parallelism in the data streams is difficult and time consuming.
In this paper we will show the suitability for scientific processing in a GPGPU environment. Given the nature of the
data streams as well as the types of processing, it will be shown that there can be a very high potential benefit in
using GPUs.
We will show several examples of how GPGPU will directly impact the current production processing, as well as
some other off-line housekeeping procedures. Also, we will demonstrate the usefulness of GPGPU programs to
accelerate in the field calculations performed by scientists in various programming environments.
Lastly, we will present a timeline for enabling GPGPU in the HIRDLS processing environment and the expected
impact on future processing of science data.
DE: 3300 ATMOSPHERIC PROCESSES
DE: 3360 Remote sensing
SC: Earth and Space Science Informatics [IN]
MN: 2007 Fall Meeting