Multiple program/multiple data molecular dynamics method with multiple time step integrator for large biological...

Area of Science:

• Computational Biology
• Biophysics
• Scientific Computing

Background:

• Parallelization of molecular dynamics (MD) simulations is crucial for studying large biological systems.
• Efficient parallel computation requires minimizing data transfer through domain decomposition.
• Optimizing the balance between real-space and reciprocal-space computations is key for handling electrostatic interactions.

Purpose of the Study:

• To introduce a novel parallelization scheme for large-scale MD simulations on CPU-only supercomputers.
• To enhance computational efficiency by optimizing data transfer and processor utilization.
• To achieve significant speedups for complex biological system simulations.

Main Methods:

• Utilized a multiple program/multiple data (MPMD) approach to separate real-space and reciprocal-space computations.
• Integrated the r-RESPA multiple time step integrator within the MPMD framework.
• Employed processor reallocation for real-space computations during skipped reciprocal-space steps.

Main Results:

• The new scheme allows the use of twice the number of processors compared to single-program approaches.
• Achieved performance of 65, 36, and 24 ns/day for 1, 8.5, and 28.8 million atom systems, respectively.
• Demonstrated speedups of 57%, 39%, and 60% for the tested systems, indicating high parallel computational efficiency.

Conclusions:

• The novel MPMD parallelization scheme offers significant speedups for large-scale MD simulations.
• This approach effectively increases processor utilization without compromising parallel efficiency.
• The method is highly suitable for large biological system investigations on massively parallel CPU architectures.