Midpoint cell method for hybrid (MPI+OpenMP) parallelization of molecular dynamics simulations.

Area of Science:

• Computational physics and chemistry
• High-performance computing
• Molecular dynamics simulations

Background:

• Molecular dynamics (MD) simulations are crucial for understanding atomic and molecular behavior.
• Efficient parallelization strategies are essential for handling large-scale MD simulations.
• Existing methods face challenges in balancing computational load and memory access.

Purpose of the Study:

• To develop a novel hybrid parallelization scheme for MD simulations.
• To improve the efficiency and scalability of MD simulations on massively parallel architectures.
• To reduce computational overhead and memory access latency.

Main Methods:

• Developed a 'midpoint cell method' combining cell-wise midpoint method with pair-wise Verlet lists.
• Implemented a hybrid Message Passing Interface (MPI) and Open Multi-Processing (OpenMP) parallelization.
• Simulation space divided into subdomains (MPI) and cells; interactions computed based on midpoint cell location.
• OpenMP threads distribute cell pairs for shared memory parallelization, optimizing memory access and reducing cache misses.

Main Results:

• Demonstrated significant parallel performance and scalability on the K computer.
• Achieved scalability up to 32,768 cores for systems with 1 million atoms.
• Calculated one MD time step for long-range interactions in 4.5 ms for a 1 million atom system with particle-mesh Ewald electrostatics.

Conclusions:

• The midpoint cell method offers an efficient and scalable parallelization strategy for MD simulations.
• The hybrid MPI+OpenMP approach effectively utilizes shared memory parallelization.
• This method significantly enhances computational performance for large-scale molecular systems.