A comparison between parallelization approaches in molecular dynamics simulations on GPUs.

Area of Science:

• Computational Chemistry
• High-Performance Computing
• Molecular Dynamics Simulations

Background:

• Molecular dynamics (MD) simulations are crucial for understanding molecular behavior.
• Efficient computation of forces is a bottleneck in MD simulations.
• Graphics Processing Units (GPUs) offer massive parallelism for accelerating computations.

Purpose of the Study:

• To compare the performance of two distinct force computation approaches for MD simulations on GPUs.
• To determine the conditions under which each approach (vertex-based vs. edge-based) is more efficient.
• To predict the future trend of these approaches with evolving GPU architectures.

Main Methods:

• Implemented and tested a "vertex-based" approach, initiating a thread per particle.
• Implemented and tested an "edge-based" approach, initiating a thread per interaction.
• Measured computation times on diverse GPU hardware, spanning different generations.

Main Results:

• The "vertex-based" approach excels in systems with numerous simple interactions per particle.
• The "edge-based" approach is superior for systems with fewer or more complex interactions.
• Performance differences were analyzed across various GPU architectures.

Conclusions:

• The "edge-based" approach is becoming increasingly competitive due to the trend of increasing GPU core counts.
• Future GPU development favoring core count over individual core power will likely favor the "edge-based" method.
• The choice of force computation method depends on system complexity and GPU architecture.