A parallel ETD algorithm for large-scale rate theory simulation

Rate theory (RT) is a commonly used method to simulate the evolution of material defects. A promising numerical method, exponential time difference (ETD), can reduce the stiff RT equations to explicit ordinary differential equations (ODEs). Previous implementations of ETD on the Sunway TaihuLight supercomputer suffer from high computation cost and poor parallel efficiency while solving a large amount of ODEs. This paper improves the algorithm with hybrid MPI+SIMD and additional instruction-level optimizations by taking advantage of the architecture of Sunway TaihuLight. The execution time of a single iteration is reduced by about 40%. Scaling from 64 to 4096 processes, the parallel efficiency of the new algorithm achieves 33.5% and 50.6% in strong and weak scalability, which corresponds to 21.4 and 32.4 in speedup, respectively.

Automated summary

This paper presents an improved algorithm that reduces the computation cost and improves parallel efficiency in solving a large amount of ODEs by leveraging the architecture of Sunway TaihuLight and employing hybrid MPI+SIMD and additional instruction-level optimizations.