Practical Performance Analysis of CPU, GPU and Xeon-Phi in Atmospheric Correction Processing for the Geostationary Ocean Color Imager (GOCI)

Abstract

In 2019, Geostationary Ocean Color Imager-II (GOCI-II) will be launched as a successor to GOCI, the worlds first geostationary ocean color sensor. The existing atmospheric correction algorithm for GOCI was designed to be processed sequentially. Also, if GOCI-II data is processed sequentially it takes more than 1 hour although GOCI-II data must be processed within 11 minutes. To meet these requirements, the parallelism was applied to the existing GOCI data processing prior to developing GOCI-II. In this paper, we have improved the processing speed of atmospheric correction algorithm using OpenMP(Open Multi-Processing) and OpenCL(Open Computing Language). The newest CPUs, Xeon-Phi and GPUs were used in experiments, and the performances of CPU- and GPU-parallelized versions were derived by comparisons with the existing sequential version. As a result, GPU version showed the performance improvements 40 times better than a sequential version and nearly twice as high as a CPU version, and its memory capacity were maintained within 3GB.ially. Also, if GOCI-II data is processed sequentially it takes more than 1 hour although GOCI-II data must be processed within 11 minutes. To meet these requirements, the parallelism was applied to the existing GOCI data processing prior to developing GOCI-II. In this paper, we have improved the processing speed of atmospheric correction algorithm using OpenMP(Open Multi-Processing) and OpenCL(Open Computing Language). The newest CPUs, Xeon-Phi and GPUs were used in experiments, and the performances of CPU- and GPU-parallelized versions were derived by comparisons with the existing sequential version. As a result, GPU version showed the performance improvements 40 times better than a sequential version and nearly twice as high as a CPU version, and its memory capacity were maintained within 3GB.