AALB: Adaptive Algorithm of Load Balance for CAM Physics on Sunway TaihuLight Supercomputer

Abstract

With expanding of parallel scales and increasing of grid precision, calculation and communication in The Community Atmosphere Model(CAM) [16] become more complicated. In this paper, various load balance algorithms, proposed by predecessors and used in CAM physics, are analyzed and tested in detail, and it is found that, improper load balance algorithm for CAM Physics may cause performance loss by 10% to 30%, and it is difficult for researchers to choose proper load balance algorithm in CAM Physics because of complex calculating and communicating mode, choosing available load balance algorithm in CAM has always been one of the most difficult and significant problem for researchers. We propose two new load balance algorithms in CAM Physics to further improve performance of CAM in special situation, one of the algorithms is global round robin splitting, which is suitable for small parallel scales, and the other is two-level regional decomposition, which is based on locality of the Community Earth System Model(CESM) [18] and the network structure of Sunway TaihuLight supercomputer, this algorithm is available for large parallel scales. At the same time, high precision model for calculating and communicating time is proposed in CAM Physics, making it easier for scientists and researchers to select proper load balance algorithm. Further more, basing on this time model, we implement an adaptive algorithm for load balance of CAM physics–AALB, AALB can select most appropriate load balance algorithm in CAM Physics adaptively. And AALB is applied in large-scale CAM on Sunway TaihuLight supercomputer platform, and successfully guides available load balance of CAM physics and improves the overall performance of CAM by 5% to 10%. To our knowledge, there is little work of load balance and time model in CAM physics with large parallel scales.