Full-neighbor-list based numerical reproducibility method for parallel molecular dynamics simulations

Abstract

The numerical nonreproducibility in parallel molecular dynamics (MD) simulations, which relates to the non-associate accumulation of float point data, leads to great challenges for development, debugging and validation. The most common solutions to this problem are using a high-precision data type or operation sorting, but these solutions are accompanied by significant computational overhead. This paper analyzes the sources of nonreproducibility in parallel MD simulations in detail. Two general solutions, namely, sorting by force component value and using an 80-bit long double data type, are implemented and evaluated in LAMMPS. To optimize the computational cost, a full-list based method with operation order sorted by particle distance is proposed, which is inspired by the spatial characteristics of MD simulations. An experiment on a system with constant energy dynamics shows that the new method can ensure reproducibility at any parallelism with an extra 50% computational overhead.