Objective: To develop a general-purpose multiple scattering radiative transfer model for atmosphere-land/ocean systems at an accuracy in the 0.1 to 2 percent range for spectral ranges from thermal infrared to ultraviolet on parallel computing architectures. In order to achieve this accuracy, the transmission function of molecular band absorption should be expressed as a sum of exponential terms for each gas species. More than a few thousand terms of the combined exponential functions may be needed for a single narrow wave number interval, thus requiring the computing power provided by high-performance parallel systems.
Approach: Two parallel algorithms for the general-purpose multiple scattering radiative transfer model ATRAD have been developed. In parallel algorithm 1, each node works independently on different wave number intervals. It is suitable for cases where the total number of the exponential terms is almost the same for each interval. Algorithm 1 provides high scalability and computing efficiency because of its perfect parallelism. In parallel algorithm 2, the nodes simultaneously work on each wave number interval so that each node has a nearly balanced computational loading. This algorithm is suitable to general cases where the total number of the exponential terms may significantly vary with wave number intervals.
Accomplishments: The first parallel algorithm has been implemented and optimized on the CRAY T3D and IBM SP-2 parallel systems. This parallel code runs 62 times faster than the sequential one and reaches 97 percent efficiency when 64 processors are used. The current level of performance is more than 6 GFLOPS on the 256-node CRAY T3D system if the number of exponential terms is the same over all wave number intervals. The second parallel algorithm was also implemented on the CRAY T3D. It can improve the performance by 50 percent over the first algorithm if the number of exponential terms varies with the wave number intervals by a factor larger than 2.
Significance: The general-purpose multiple scattering radiative transfer model for atmosphere-land/ocean systems can be applied more efficiently to applications in the global change research program by using high-performance parallel computing systems.
Status/Plans: The parallel algorithms will be ported onto different parallel systems, including massively parallel supercomputers and clusters of UNIX workstations.
Point of Contact:
Zhengming Wan
University of California, Santa Barbara
wan@icess.ucsb.edu
805-893-4541