[Paper] Strategies for Molecular Dynamics using Hybrid Systems: LAMMPS Use Case

Source: arXiv - 2606.02319v1

Overview

The paper evaluates how the popular molecular‑dynamics engine LAMMPS performs when simulating a coarse‑grained antimicrobial peptide (Tritrpticin) on modern high‑performance‑computing (HPC) clusters. By comparing pure MPI against hybrid MPI + OpenMP runs across up to 1024 cores, the authors expose the scaling limits of classic MPI‑only parallelism and demonstrate why mixed‑mode execution is becoming the preferred strategy for large‑scale biomolecular workloads.

Key Contributions

• Systematic scalability study of LAMMPS for coarse‑grained biomolecular simulations on up to 8 nodes (1024 cores).
• Head‑to‑head comparison of pure‑MPI vs. hybrid MPI + OpenMP configurations, including detailed timing, speed‑up, and parallel‑efficiency metrics.
• Profiling of internal LAMMPS phases, pinpointing communication and electrostatic calculations as the dominant bottlenecks at high MPI counts.
• Guidelines for optimal parallel granularity, showing how hybrid execution better matches the NUMA memory hierarchy of contemporary many‑core CPUs.
• Open‑source reproducibility package (scripts, input files, and performance logs) that can be reused for other biomolecular systems.

Methodology

1. Workload – The authors built a coarse‑grained model of the peptide Tritrpticin (PDB 1D6X) using the MARTINI force field, a standard choice for speeding up biomolecular dynamics while preserving essential physics.
2. Hardware – Tests were run on a typical HPC cluster: each node hosts multiple sockets with 16‑core Intel Xeon CPUs, connected via InfiniBand. The node count varied from 1 to 8, giving 64–1024 logical cores.
3. Parallel configurations
   • Pure MPI: one MPI rank per core (up to 1024 ranks).
   • Hybrid MPI + OpenMP: a small number of MPI ranks per node (e.g., 2‑4) with multiple OpenMP threads per rank.
4. Metrics collected – wall‑clock time, speed‑up relative to a single‑core baseline, parallel efficiency, statistical variability (standard deviation across runs), and a breakdown of LAMMPS’ internal timers (force computation, neighbor list, communication, etc.).
5. Analysis – The authors plotted scaling curves, computed efficiency loss, and correlated performance drops with specific LAMMPS modules (e.g., electrostatics, MPI_Allreduce).

Results & Findings

Key observations

• Pure MPI scales well on a single node (up to 64 cores) but hits a steep efficiency wall beyond that due to inter‑node communication latency and synchronization overhead.
• Hybrid runs keep the same or slightly better wall‑time at 512–1024 cores while using far fewer MPI ranks, which cuts down on collective communication traffic.
• Communication and electrostatic routines dominate the runtime at the highest pure‑MPI scale (≈ 45 % of total time), confirming that the cost of moving data outweighs raw compute.
• NUMA‑aware thread placement in the hybrid mode improves memory bandwidth utilization, especially for the short‑range neighbor‑list updates.

Practical Implications

• For developers of MD‑based pipelines (e.g., drug‑discovery or protein‑design workflows), the study suggests configuring LAMMPS with a modest number of MPI ranks per node (2‑4) and leveraging OpenMP threads to saturate the cores. This yields faster turnaround without rewriting simulation scripts.
• HPC administrators can allocate resources more efficiently: instead of packing a node with 64 MPI tasks, they can schedule hybrid jobs that leave network bandwidth free for other users, improving overall cluster utilization.
• Software engineers building wrappers or GUIs around LAMMPS should expose hybrid‑mode options by default, perhaps auto‑detecting the node’s core count and NUMA topology.
• Portability to cloud‑based HPC – many cloud instances expose many‑core VMs with shared memory; hybrid MPI + OpenMP aligns naturally with these environments, reducing the need for costly inter‑VM networking.
• Future extensions (e.g., coupling LAMMPS with machine‑learning potentials) will inherit the same communication patterns, so the hybrid strategy will likely remain beneficial.

Limitations & Future Work

• The analysis focuses on a single coarse‑grained peptide system; results may differ for larger, more heterogeneous biomolecular assemblies or all‑atom simulations.
• Only Intel Xeon CPUs and InfiniBand interconnects were examined; performance on AMD EPYC, ARM, or GPU‑accelerated nodes remains an open question.
• The study does not explore dynamic load‑balancing or task‑based runtimes (e.g., using MPI+Thread or OpenMP + MPI‑3 RMA), which could further mitigate communication bottlenecks.
• Authors propose extending the benchmark suite to include GPU‑offloaded force kernels and to evaluate energy‑efficiency metrics (power‑to‑solution) for greener HPC operation.

Authors

• Paulo Henrique Leme Ramalho
• Dennis Alves Pedersen
• Fábio Andrijauskas

Paper Information

• arXiv ID: 2606.02319v1
• Categories: cs.DC
• Published: June 1, 2026
• PDF: Download PDF