[Paper Review] Long range forces in a performance portable Molecular Dynamics framework

William R. Saunders, James P. Grant+1 more|arXiv (Cornell University)|Aug 3, 2017

Parallel Computing and Optimization Techniques3 citations

TL;DR

This paper presents a performance-portable molecular dynamics framework that integrates Ewald summation for long-range electrostatic forces using a domain-specific language (DSL) and code generation. By combining MPI+OpenMP parallelism and automatic optimization, the framework achieves strong scalability and competitive performance—especially on moderate core counts—demonstrating efficient portability across modern manycore architectures while supporting complex electrostatic interactions critical for accurate MD simulations.

ABSTRACT

Molecular Dynamics (MD) codes predict the fundamental properties of matter by following the trajectories of a collection of interacting model particles. To exploit diverse modern manycore hardware, efficient codes must use all available parallelism. At the same time they need to be portable and easily extendible by the domain specialist (physicist/chemist) without detailed knowledge of this hardware. To address this challenge, we recently described a new Domain Specific Language (DSL) for the development of performance portable MD codes based on a "Separation of Concerns": a Python framework automatically generates efficient parallel code for a range of target architectures. Electrostatic interactions between charged particles are important in many physical systems and often dominate the runtime. Here we discuss the inclusion of long-range interaction algorithms in our code generation framework. These algorithms require global communications and careful consideration has to be given to any impact on parallel scalability. We implemented an Ewald summation algorithm for electrostatic forces, present scaling comparisons for different system sizes and compare to the performance of existing codes. We also report on further performance optimisations delivered with OpenMP shared memory parallelism.

Motivation & Objective

• Address the challenge of implementing efficient, portable long-range electrostatic forces in molecular dynamics simulations without requiring low-level parallel programming expertise.
• Enable domain scientists (chemists/physicists) to express complex force kernels at a high level while achieving high performance across diverse hardware.
• Extend the performance-portable MD framework from [3] to support long-range interactions, particularly electrostatics via Ewald summation.
• Improve parallel scalability for large-scale simulations by introducing a hybrid MPI+OpenMP backend for better load balancing and reduced communication overhead.
• Achieve performance comparable to established MD codes like DL_POLY_4 while maintaining code portability and extensibility.

Proposed method

• Use a domain-specific language (DSL) built on Python to express particle interaction kernels at a high level, abstracting hardware-specific details.
• Apply a code generation system to automatically produce optimized, parallel C code for multiple architectures, including multi-CPU and multi-GPU systems.
• Implement the Ewald summation algorithm to compute long-range electrostatic forces by splitting the potential into real-space and Fourier-space contributions.
• Use a hybrid MPI+OpenMP parallelization strategy: MPI for inter-node communication and OpenMP for shared-memory parallelism within nodes, reducing communication costs.
• Generate thread-safe reduction code for global array operations to handle write conflicts in parallel execution without requiring user intervention.
• Optimize performance via vectorization (e.g., Intel compiler-generated packed instructions) and tuning of Ewald parameters (α and rc) to balance real- and Fourier-space work.

Experimental results

Research questions

• RQ1 Can a performance-portable MD framework efficiently and portably implement long-range electrostatic forces using Ewald summation?
• RQ2 How does the performance of the Ewald implementation compare to established MD codes like DL_POLY_4 in terms of absolute runtime and scalability?
• RQ3 To what extent does the hybrid MPI+OpenMP backend improve strong scaling compared to pure MPI, especially for large core counts?
• RQ4 What is the impact of non-optimized Ewald parameters (α, rc) on performance when scaling beyond one node?
• RQ5 Can the framework maintain high performance while supporting complex, long-range interactions without sacrificing portability or ease of use for domain scientists?

Key findings

• The Ewald summation implementation achieves O(N^3/2) computational complexity, confirming theoretical expectations, with 87% of runtime spent on electrostatics for N = 1.8×10^5 particles.
• On a single 8-core CPU, the framework’s performance is comparable to DL_POLY_4 for small systems, but DL_POLY_4 outperforms it by an order of magnitude for larger systems due to its SPME method.
• The hybrid MPI+OpenMP backend enables strong scaling to 256 cores (16 nodes), achieving good parallel efficiency and overcoming the scalability limits of pure MPI.
• The MPI+OpenMP implementation on one node achieved 34% of peak floating-point vector performance, with the Fourier loop being the most computationally expensive component.
• Vectorization of the Fourier-space loop over k-space quadrants was successfully enabled by the Intel compiler, confirming effective use of SIMD instructions.
• The framework maintains good performance and scalability even with suboptimal Ewald parameters (α = 0.032, rc = 19Å), with minimal impact on single-node performance compared to optimal settings.