Molecular dynamics (MD) simulations are playing an increasingly important role in several research areas. The most frequently used potentials in MD simulations are pair-wise potentials. Due to the memory wall, computing pair-wise potentials on many-core processors is usually memory bounded.
In this paper, we take the SW26010 processor as an example platform to explore the possibility of breaking the memory bottleneck by improving data reuse via neighbor-list-free methods. We use cell-lists instead of neighbor-lists in the potential computation, and apply several novel optimization methods. These methods include; an adaptive replica arrangement strategy, a parameter-profile data structure and a particle-cell cutoff checking filter. Also, an incremental cell-list building method is realized to accelerate the construction of cell-lists.
Experiments show our implementation is up to 170% faster than previous ports on the same platform. And our ESMD framework can scale to 1,024 nodes with a weak scalability of 95%.