Heterogeneous parallelization and acceleration of molecular dynamics simulations in GROMACS

Area of Science:

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

Background:

• Graphics processing units (GPUs) offer significant performance gains for molecular dynamics (MD) simulations.
• Adapting fundamental MD algorithms (e.g., Verlet lists, pair searching) is crucial for leveraging accelerator hardware.
• The GROMACS codebase has been a focus for integrating and optimizing these hardware accelerations.

Purpose of the Study:

• To present the heterogeneous parallelization and acceleration design for molecular dynamics within the GROMACS codebase.
• To detail algorithmic reformulations enabling efficient utilization of both GPUs and CPUs.
• To showcase performance improvements from single GPU to multi-node parallelization.

Main Methods:

• Developed a general cluster-based approach for pair lists and non-bonded interactions.
• Implemented heterogeneous parallelization utilizing GPU and CPU single instruction, multiple data (SIMD) acceleration.
• Introduced dual pair lists with rolling pruning updates for efficient accelerator use.
• Enabled direct GPU-GPU communication and integrated GPU-specific optimizations.

Main Results:

• Achieved efficient load balancing of tasks between CPUs and GPUs.
• Tuned algorithm work efficiency for optimal performance on different hardware types.
• Demonstrated excellent performance scaling from single GPU to multi-GPU and multi-node configurations.
• Validated the effectiveness of dual pair lists and GPU-GPU communication.

Conclusions:

• The presented heterogeneous parallelization design significantly enhances molecular dynamics simulation performance in GROMACS.
• Efficiently combining CPU and GPU resources with optimized algorithms unlocks substantial computational gains.
• The GROMACS enhancements provide a scalable and efficient platform for large-scale molecular dynamics studies.