Performance Portable Electrostatics for Exascale Atomistic Simulations [14:50 - 15:10]
Authors: Rodrigo Bartolomeu, René Halver, Godehard Sutmann
Long‑range Coulomb interactions remain a principal bottleneck in particle‑based atomistic simulations, especially when periodic boundary conditions are required. Mesh‑based Ewald approaches such as the particle‑particle particle‑mesh (P3M) method reduce the computational complexity to but introduce severe challenges for parallel scalability, communication overhead, and efficient exploitation of heterogeneous CPU/GPU architectures. Existing implementations are tightly coupled to monolithic molecular‑dynamics packages, limiting their reuse in emerging high‑performance‑computing (HPC) workflows.
We introduce an exascale‑ready, performance‑portable library that delivers unified implementations of both the classical Ewald summation and the P3M method. The core solvers are written using the Kokkos programming model, allowing a single source base to be compiled for CPUs and GPUs from multiple vendors while delegating execution and memory management to portable abstractions. To further enhance portability and efficiency, it integrates the Cabana library for cache‑friendly particle data layouts and communication‑aware operations, as well as the HeFFTe library for architecture‑aware, optimized FFTs required by the reciprocal‑space part of the mesh-based solvers.
A key design principle is a modular execution model that separates short‑range (real‑space) and long‑range (reciprocal‑space) electrostatics into distinct computational partitions. On modern HPC systems, this enables independent allocation of CPU cores, GPUs and memory according to each partition's needs, improving load-balancing and allowing for heterogeneous job partitioning.
We present strong‑scaling benchmarks on a range of platforms (multi‑core CPUs, NVIDIA and AMD GPUs, and mixed CPU‑GPU nodes). For small problem sizes the performance is limited by communication latency, whereas larger systems achieve near‑ideal scaling across dozens of GPUs, confirming that the FFT‑driven long‑range component can be efficiently parallelised when compute‑communication balance is respected. Across all tested architectures the methods maintains numerical accuracy comparable to reference implementations while delivering speed‑up over the sequential execution mode.
Results demonstrate that a portable, decoupled library can provide competitive, scalable long‑range electrostatics for atomistic simulations on current and future exascale systems. Our library thus offers a flexible foundation for integrating high‑performance electrostatic solvers into diverse simulation workflows, accelerating the scientific discovery in materials science and soft‑matter physics.