Load-balanced diffusion Monte Carlo method with lattice regularization

Area of Science:

• Computational Physics
• Quantum Chemistry
• High-Performance Computing

Background:

• Ab initio quantum Monte Carlo (QMC) offers a stochastic solution to the many-body Schrödinger equation without one-body approximations.
• Diffusion Monte Carlo (DMC), a QMC variant, reliably projects onto the ground state using the fixed-node approximation.
• Lattice regularized diffusion Monte Carlo (LRDMC) is a practical DMC implementation but suffers from load imbalance issues in parallelization.

Purpose of the Study:

• To develop and present a novel load-balanced LRDMC algorithm.
• To address and mitigate the inherent load imbalance in conventional LRDMC.
• To significantly improve weak-scaling parallel efficiency on modern architectures.

Main Methods:

• Implementation of a new LRDMC algorithm designed for inherent load balancing.
• Utilized ensembles of walkers (Nw) for parallelization on high-performance computing systems.
• Tested the algorithm using binding energy calculations for a water-methane complex on the Leonardo supercomputer with NVIDIA A100 GPUs.

Main Results:

• The load-balanced LRDMC algorithm yields results consistent with the conventional method.
• Achieved high parallel efficiency (∼98%) up to 512 GPUs (Nw = 51,200).
• Demonstrated a speedup of ×1.24 compared to conventional LRDMC with the same walker count, and ×1.18 even with reduced walkers (Nw = 400).

Conclusions:

• The developed load-balanced LRDMC algorithm effectively resolves load imbalance issues in parallel QMC simulations.
• This advancement leads to superior hardware utilization and parallel efficiency on large-scale computing resources.
• The method shows significant potential for accelerating complex quantum mechanical calculations in computational physics and chemistry.