On-the-fly clustering for exascale molecular dynamics simulations.

To achieve approximately 160 × 10^3 atoms per core, we note that while the scalability of our existing blue system is not ideal, the scalability of molecular dynamics is nearly perfect, and communication issues are minimal. Although our analysis is less efficient, the bottleneck we identified indicates that reducing the information from a single molecular interaction (MI) is manageable.

We aim to ensure that our analysis does not require excessive computational time. It is critical to consider the time associated with molecular dynamics; specifically, the analysis must not exceed the computational cost of a single step of dynamics. The following graph illustrates the ratio of analysis time to one step of molecular dynamics across varying loads and core counts.

The light blue points represent data from low load strength scaling, while the pink points correspond to high loads. The points with varying colors reflect weak scaling. As observed, the load increases with the number of cores. When optimization is adequately implemented, the analysis time approaches 5% of the molecular dynamics time, which is satisfactory. Consequently, this analysis will not be conducted after every step but rather every 10 or 100 steps, making its impact negligible compared to the overall cost of molecular dynamics.