Consideration of data load time on modern processors for the Verlet table and linked-cell algorithms

Area of Science:

• Computational physics
• Molecular dynamics simulations
• Algorithm analysis

Background:

• Neighbor search algorithms are crucial for calculating interatomic potentials in molecular dynamics.
• Verlet table (VT) and linked-cell (LC) methods are commonly used for neighbor searching.
• VT is often considered more efficient than LC.

Purpose of the Study:

• To analyze the efficiency of VT and LC neighbor search algorithms.
• To determine the optimal algorithm for molecular dynamics simulations under various conditions.
• To investigate performance improvements through algorithm modifications and hardware utilization.

Main Methods:

• Comparative analysis of VT and LC algorithms based on memory data load operations.
• Mathematical condition derived for LC's advantage: (4π/6)ρr(skin)3/27 ≫ 1.
• Implementation and testing within MOLKERN simulation software using SIMD and multithreading.

Main Results:

• LC demonstrates significantly fewer memory data load operations than VT at high particle densities.
• A newly developed LC-based code outperformed the VT by a factor of almost 2.
• Modified VT algorithms improved performance but did not surpass LC's efficiency.

Conclusions:

• The linked-cell method is superior to the Verlet table method for molecular dynamics simulations when particle density is high.
• Optimizing memory access is critical for high-performance computing in simulations.
• LC offers a more efficient approach for neighbor searching in dense systems.