Performance Modeling, Analysis, and Optimization of Cell-List Based Molecular Dynamics

We have developed a performance prediction model for non-bonded interaction computations in molecular dynamics simulations, thereby predicting the optimal cell dimension in a linked-list cell method. The model expresses computation time in terms of the number and unit computation time of key operations. The model accurately estimates the number of operations during the simulations with the maximum standard error of 10.6% compared with actual measurements. Then, the unit computation times of the operations are obtained by bisquare regression. Analysis of this model reveals that the optimal cell dimension to minimize the computation time is determined by a trade-off between decreasing search space and increasing linked-list cell access for smaller cells. The model predicts the optimal cell dimension correctly for 80% of all tested cases, resulting in an average speedup of 10% and 52% for the cutoff radius of interaction, 6.6 and 10.0 Å, respectively.