Porting and Optimizing VASP on the SW26010

Abstract

VASP (Vienna Ab initio Simulation Package) is a prevalent first-principle software framework. It is so widely used that its runtime usually dominates the usage of current supercomputers. The porting and optimization of VASP to the Sunway TaihuLight supercomputer, a newly heterogeneous many-core platform based on SW26010 CPU, becomes of great importance. In this paper, we focus on the challenges in porting and optimizing VASP on the SW26010 CPU. Optimizations on three types of time-consuming kernels, which include matrix operations, FFT, and certain domain-specific computing primitives, are carried out base on thorough performance profiling. The experimental results are shown by the case of RELAX, where speedup of 2.90x and 4.48x is sustained respectively for both of the iterative diagonalization methods in VASP, RMM-DIIS (RMM) and block Davidson (DAV).