Algorithm to minimize MPI communications in the parallelized fast multipole method combined with molecular dynamics...

Area of Science:

• Computational physics and chemistry
• High-performance computing
• Materials science

Background:

• Exascale supercomputers enable unprecedented scale in scientific simulations.
• Molecular dynamics (MD) calculations require efficient parallelization for large-scale, long-time simulations.
• The Fast Multipole Method (FMM) is crucial for calculating electrostatic interactions in MD.

Purpose of the Study:

• To develop a novel algorithm for improving Message Passing Interface (MPI) communication performance in MPI-parallelized FMM-MD calculations.
• To reduce the volume of communication data, specifically atomic coordinates and multipole coefficients.
• To enhance the efficiency of large-scale and long-time MD simulations on exascale systems.

Main Methods:

• Proposed a new algorithm, the minimum-transferred data (MTD) method.
• Applied the MTD method to MPI-parallelized FMM combined with MD under 3D periodic boundary conditions.
• Focused on optimizing communication of atomic coordinates and multipole coefficients.

Main Results:

• The MTD method drastically reduces communication data required for electrostatic interaction calculations.
• The communication data reduction rate increases with the number of FMM levels and MPI processes.
• For very large systems, the reduction rate can exceed 50% with increased MPI processes.

Conclusions:

• The MTD method significantly improves the efficiency of FMM-MD calculations.
• This algorithm is crucial for achieving efficient massive MPI-parallelization on exascale supercomputers.
• Enables large-scale and long-time MD simulations crucial for scientific breakthroughs.