混合粒子云代码dHybridR的性能可移植GPU加速

中文摘要 AI 辅助

混合粒子云模拟被广泛用于研究无碰撞天体物理和空间等离子体中的动力学过程，然而大规模三维运行的高计算成本在很大程度上将生产性研究限制在二维或受限域内。为了应对这一挑战，我们提出了混合粒子云代码dHybridR的一种性能可移植GPU实现。该实现将OpenMP目标卸载与针对计算最昂贵操作的专用SYCL内核相结合，同时保留了一个统一的CPU-GPU代码库，支持Intel、AMD和NVIDIA GPU。在百亿亿次超级计算机Aurora和Frontier上，dHybridR在49,152个加速器上实现了86%至97%的弱扩展效率，并且在每单元256个粒子的情况下，其全节点GPU吞吐量高达向量化CPU实现的32倍，每次粒子更新的能耗约降低90%。据我们所知，没有其他混合粒子云代码在此规模上报告过GPU性能，这使得dHybridR在利用百亿亿次系统方面具有独特优势。这些进展大大降低了大规模三维混合动力学无碰撞等离子体模拟的计算障碍。

英文摘要

Hybrid particle-in-cell simulations are widely used to study kinetic processes in collisionless astrophysical and space plasmas, yet the high computational cost of large-scale three-dimensional runs has largely confined production studies to two dimensions or restricted domains. To address this challenge, we present a performance-portable GPU implementation of the hybrid particle-in-cell code dHybridR. The implementation combines OpenMP target offloading with specialized SYCL kernels for the most computationally expensive operations, while preserving a unified CPU-GPU codebase that supports Intel, AMD, and NVIDIA GPUs. On the exascale supercomputers Aurora and Frontier, dHybridR achieves weak scaling efficiencies of $86\%$ to $97\%$ across 49,152 accelerators, and at 256 particles per cell, its full-node GPU throughput is up to $32\times$ that of the vectorized CPU implementation at approximately $90\%$ less energy per particle-update. To our knowledge, no other hybrid particle-in-cell code has reported GPU performance at this scale, leaving dHybridR uniquely positioned to exploit exascale systems. These advances substantially lower the computational barrier to large-scale three-dimensional hybrid-kinetic simulations of collisionless plasmas.

发表机构

• The University of Chicago（芝加哥大学）
• Cerebras Systems（Cerebras系统）
• Enrico Fermi Institute, The University of Chicago（芝加哥大学恩里科·费米研究所）

机构由 AI 辅助整理，请以论文原文为准。