A parallel CUDA implementation of the Gauss–Legendre–spherical- <i>t</i> method for electrostatic interactions

Abstract

Computing electrostatic interactions remains the bottleneck of molecular dynamics (MD) simulations despite more than a century of effort in developing methods to accelerate the calculation. Previously, we have developed the spherical grids and treecode and Gauss–Legendre–spherical-t (GLST) algorithms for electrostatic interactions. Here, we explain the computational details and discuss the performance of GLST. The GLST algorithm achieves O(N) scaling and should be less demanding in parallel communication compared with the widely used particle mesh Ewald method and likely comparable to the communication costs of the fast multipole method. We find that GLST is suitable for rapid calculation of long-range electrostatic interactions in MD simulations as it has highly tunable accuracy and should scale well on massively parallel computing architectures. The GLST software presented here is available as a standalone library on GitHub.