模拟多发动机航天器：通过信息几何正则化实现超过10^15自由度的计算
计算物理
2025-12-01 v4 计算工程、金融与科学
流体动力学
摘要
我们 presented an optimized implementation of the recently proposed information geometric regularization (IGR) for unprecedented scale simulation of compressible fluid flows applied to multi-engine spacecraft boosters. 我们在计算成本、内存占用和能耗效率等指标上均优于当前最先进的计算流体动力学(CFD)技术。统一的CPU--GPU或APU平台内存架构使问题规模几乎无额外开销地获得提升。采用混合半精度/单精度存储与计算，适用于良好条件的数值场景。我们模拟了200万亿个网格点和10^15个自由度的流场，超越了当前纪录的20倍。相较于优化基准，实现了4倍的单位时间加速。在OLCF Frontier、LLNL El Capitan和CSCS Alps上使用完整系统均实现了理想的弱缩放。极端条件下的强缩放也接近理想，在CSCS Alps上以8节点基准为例实现80%效率，并可扩展至整个系统。
引用
@article{arxiv.2505.07392,
title = {Simulating many-engine spacecraft: Exceeding 1 quadrillion degrees of freedom via information geometric regularization},
author = {Benjamin Wilfong and Anand Radhakrishnan and Henry Le Berre and Daniel J. Vickers and Tanush Prathi and Nikolaos Tselepidis and Benedikt Dorschner and Reuben Budiardja and Brian Cornille and Stephen Abbott and Florian Schäfer and Spencer H. Bryngelson},
journal= {arXiv preprint arXiv:2505.07392},
year = {2025}
}
备注
11 pages, 8 figures, 4 tables. SC25 ACM Gordon Bell Prize Finalist