Modeling the Parallel Efficiency of Density Functional Theory Based Jobs on Sunway TaihuLight

Existing supercomputers have a lot of computing resources, the scalable parallel efficiency of supercomputers can significantly affect the wall time and core hours of jobs on supercomputers. In this paper, we take density functional theory based VASP jobs as one kind of typical applications to model the parallel efficiency of the Sunway TaihuLight, one of the supercomputers among the world. We investigate not only the different number of cores, but also different parallel configurations with the same number cores that can effect the performance. A heuristic algorithm is provided to optimize the parallel performance quickly. We find that the parallel efficiency of Sunway TaihuLight can be perfectly modeled with a Logarithmic function. Experimental results show that the parallel efficiency of VASP jobs on Sunway TaihuLight can achieve up to 80% when we double the number of cores. The proposed method can be employed to optimize the job performance on Sunway TaihuLight.