Performance improvements of an atmospheric radiative transfer model on GPU-based platform using CUDA

Classical applications of Atmospheric Radiative Transfer Model (ARTM) for modelization of absorption coefficient line-by-line on the atmosphere consume large computational time since seconds up to a few minutes depending on the atmospheric characterization chosen. ARTM is used together with Ground- Based or Satellite measurements to retrieve atmospheric parameters such as ozone, water vapour and temperature profiles. Nowadays in the Atmospheric Observatory of Southern Patagonia (OAPA) at the Patagonian City of Río Gallegos have been deployed a Spectral Millimeter Wave Radiometer belonging Nagoya Univ. (Japan) with the aim of retrieve stratospheric ozone profiles between 20-80 Km. Around 2 GBytes of data are recorder by the instrument per day and the ozone profiles are retrieving using one hour integration spectral data, resulting at 24 profiles per day. Actually the data reduction is performed by Laser and Application Research Center (CEILAP) group using the Matlab package ARTS/QPACK2. Using the classical data reduction procedure, the computational time estimated per profile is between 4-5 minutes determined mainly by the computational time of the ARTM and matrix operations. We propose in this work first add a novel scheme to accelerate the processing speed of the ARTM using the powerful multi-threading setup of GPGPU based at Compute Unified Device Architecture (CUDA) and compare it with the existing schemes. Performance of the ARTM has been calculated using various settings applied on a NVIDIA graphic Card GeForce GTX 560 Compute Capability 2.1. Comparison of the execution time between sequential mode, Open-MP and CUDA has been tested in this paper.XV Workshop de Procesamiento Distribuido y Paralelo (WPDP)Red de Universidades con Carreras en Informática (RedUNCI