CAREER: HiPer: A CFD solver for High-Performance Turbulent Flow Simulations on Massively Parallel Machines

职业：HiPer：用于大规模并行机器上高性能湍流模拟的 CFD 求解器

• 批准号: 1750549
• 负责人: Aparna Chandramowlishwaran
• 金额: $ 50万
• 依托单位: University of California-Irvine
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2018
• 资助国家: 美国
• 起止时间: 2018-03-01 至 2024-05-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1750549&HistoricalAwards=false
• 关键词: CAREER; HiPer; CFD; solver; Performance

Evolution of physics-based simulations and Computational Fluid Dynamics (CFD) in particular has fundamentally reshaped the design and engineering process in the last several decades. However, in spite of noteworthy success, today's CFD still remains limited to a small design space. One of the grand challenge problems is the simulation of a full aircraft envelope. Today's computing platforms are based on massive parallelism and heterogeneous processor designs to deliver petascale performance (10^15 floating point operations per second). This research has the potential to effectively harness this level of performance and significantly address grand challenge problems through advances in numerical schemes, efficient parallel algorithms, and implementation strategies to enable underlying simulations that are currently infeasible. It also aims at transforming CFD simulation capabilities to dramatically reduce the cost and time needed to solve complex multi-physics problems. As such, this research fills a gap in the understanding that lies at the intersection of computational fluid dynamics and parallel computing. This interdisciplinary research also shapes the next-generation of students and researchers with a multi-faceted skill set required to solve challenging problems at the boundary of domain sciences, applied mathematics, and computer science to promote the progress of science and advance national prosperity and welfare as stated by NSF's mission.This research creates HiPer, a CFD solver for high-performance turbulent flow simulations on massively parallel machines. HiPer solves the Navier-Stokes equations on multi-block structured grids for complex geometries by combining the following key components: (i) a novel time-delayed implicit time-marching scheme tailored for heterogeneous architectures; (ii) parallelization strategies for shared- and distributed-memory systems aimed at reducing the synchronization and communication time; and a hybrid multi-block structured grid enhanced by geometric multigrid to increase convergence as well as reduce the number of grid cells. Applications that are drivers in the near- and long-term serve as benchmarks for continually measuring progress towards the grand challenge goals. In particular, two application case studies serve as drivers for the near-term: (a) the simulation of the complex unsteady flow through multi-stage compressors and turbines, and (b) noise generation and propagation in a high-speed turbulent jet. The intellectual merit of this work is the development of novel numerical techniques, parallelization strategies, and scalable software that enable turbulent-separated flow simulations that are computationally intractable today. To engage and inspire young generations to this approach, this project strives to (i) organize hands-on workshops at relevant conferences, (ii) design and develop an educational kit targeted to teaching CFD and high-performance computing (HPC) concepts, (iii) design lab-based courses on HPC for computational scientists, including hands-on labs using HiPer on large-scale systems for on-site students, and (iv) release videos and guest lectures through University of California's Early Academic Outreach Program. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

在过去的几十年里，基于物理的模拟和计算流体动力学（CFD）的发展从根本上重塑了设计和工程过程。 然而，尽管取得了显着的成功，当今的CFD仍然局限于较小的设计空间。 最大的挑战问题之一是全飞机包线的模拟。 当今的计算平台基于大规模并行和异构处理器设计，以提供千万亿次性能（每秒10^15次浮点运算）。 这项研究有可能有效地利用这种水平的性能，并通过数值方案，高效的并行算法和实施策略的进步，使底层模拟，目前是不可行的，显着解决重大挑战的问题。 它还旨在转换CFD模拟功能，以大幅降低解决复杂多物理场问题所需的成本和时间。 因此，这项研究填补了计算流体动力学和并行计算交叉领域的理解空白。 该跨学科研究还塑造了下一代的学生和研究人员，他们需要多方面的技能来解决领域科学，应用数学和计算机科学的边界上的挑战性问题，以促进科学的进步，促进国家的繁荣和福利，正如NSF的使命所述。本研究创建了HiPer，一个CFD求解器，用于大规模并行机器上的高性能湍流模拟。 HiPer通过结合以下关键组件在复杂几何的多块结构网格上求解Navier-Stokes方程：（i）一种针对异构体系结构的新型时间延迟隐式时间推进方案;（ii）针对共享和分布式存储系统的并行化策略，旨在减少同步和通信时间;采用几何多重网格增强的混合多块结构网格，以提高收敛性并减少网格单元数。 作为近期和长期驱动力的应用程序可作为不断衡量实现重大挑战目标的进展的基准。 特别是，两个应用案例研究作为近期的驱动力：（a）通过多级压缩机和涡轮的复杂非定常流的模拟，以及（B）高速湍流射流中的噪声产生和传播。 这项工作的智力价值是开发新的数值技术，并行化策略和可扩展的软件，使并行分离流模拟，今天是计算上棘手的。 为了吸引和激励年轻一代采用这种方法，该项目努力（i）在相关会议上组织实践研讨会，（ii）设计和开发针对CFD和高性能计算（HPC）概念教学的教育工具包，（iii）为计算科学家设计基于实验室的HPC课程，包括现场学生在大规模系统上使用HiPer的实践实验室，以及（iv）通过加州大学的早期学术推广计划发布视频和客座讲座。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

SURFNet: Super-Resolution of Turbulent Flows with Transfer Learning using Small Datasets

• DOI: 10.1109/pact52795.2021.00031
• 发表时间: 2021-08
• 期刊: 2021 30th International Conference on Parallel Architectures and Compilation Techniques (PACT)
• 作者: Octavi Obiols-Sales;Abhinav Vishnu;Nicholas Malaya;Aparna Chandramowlishwaran

Roofline Guided Design and Analysis of a Multi-stencil CFD Solver for Multicore Performance

• DOI: 10.1109/ipdps.2018.00085
• 发表时间: 2018-05
• 期刊: 2018 IEEE International Parallel and Distributed Processing Symposium (IPDPS)
• 作者: Bahareh Mostafazadeh;F. Marti;Feng Liu;Aparna Chandramowlishwaran

Mosaic flows: A transferable deep learning framework for solving PDEs on unseen domains

• DOI: 10.1016/j.cma.2021.114424
• 发表时间: 2021-04
• 期刊: Computer Methods in Applied Mechanics and Engineering
• 影响因子: 7.2
• 作者: Hengjie Wang;R. Planas;Aparna Chandramowlishwaran;R. Bostanabad

CFDNet: a deep learning-based accelerator for fluid simulations

• DOI: 10.1145/3392717.3392772
• 发表时间: 2020-05
• 期刊: Proceedings of the 34th ACM International Conference on Supercomputing
• 作者: Octavi Obiols-Sales;Abhinav Vishnu;Nicholas Malaya;Aparna Chandramowlishwaran

Pencil: A Pipelined Algorithm for Distributed Stencils

• DOI: 10.1109/sc41405.2020.00089
• 发表时间: 2020-11
• 期刊: SC20: International Conference for High Performance Computing, Networking, Storage and Analysis
• 作者: Hengjie Wang;Aparna Chandramowlishwaran