ddcMD: A fully GPU-accelerated molecular dynamics program for the Martini force field

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Scientific Computing

Background:

• Molecular dynamics simulations are crucial for understanding molecular behavior.
• The Martini force field is widely used for coarse-grained simulations.
• Accelerating simulations is essential for tackling larger and more complex systems.

Purpose of the Study:

• To implement the Martini force field and simulation integration steps on graphics processing units (GPUs).
• To enhance the performance of the ddcMD molecular dynamics program.
• To enable faster and more efficient molecular simulations.

Main Methods:

• Porting the Martini force field to a heterogeneous programming model for GPU acceleration.
• Implementing the entire integration step (thermostat, barostat, constraint solver) on GPUs.
• Benchmarking the ddcMD program against GROMACS for Martini simulations.

Main Results:

• Achieved a 278-fold speedup using one GPU compared to one CPU core.
• Demonstrated performance of 1.04 µs/day on one NVIDIA V100 GPU for a 136k particle system.
• Attained an aggregate performance of 6.19 µs/day on one Summit node with six GPUs.
• Offloaded all computations to the GPU, requiring only one CPU core for I/O management.

Conclusions:

• The GPU implementation in ddcMD significantly accelerates molecular dynamics simulations.
• This advancement allows for more complex simulations and frees up CPU resources.
• The open-source ddcMD code is available for broader scientific use.