Computational Fluid Dynamics in the Exascale Era of Computation

自然界中到处都是液体。计算流体动力学(CFD)是利用计算机获得描述流体流动的方程的解的广泛领域。使用CFD技术进行流体流动模拟已经成为研究一系列科学学科问题的不可或缺的工具。模拟使我们能够验证我们对复杂Pemonena的理解，在许多情况下，它提供了从数学或实验角度研究难以解决的问题的唯一潜在途径。模拟的范围不断扩大，反映了将小规模过程与大规模特征连接起来的需求，以及创建涵盖所有相关物理过程的完全一致的模拟的动力。这种模拟的规模和复杂性是允许的，因为超级计算集群的计算能力随着时间的推移继续呈指数级增长。在接下来的十年内，第一批超级计算集群将上线，其计算速度为每秒10^18次浮点运算(Exaflop)。然而，在设计能够扩展百亿级集群的CFD解算器的并行算法方面存在着巨大的挑战。这些集群将拥有更强的计算能力，而不是通过单个CPU变得更快，而是通过纯粹的CPU数量，以及通过包括CPU和GPU(图形处理单元)的混合。通信负载和内存瓶颈将需要克服，而如今罕见到可以忽略的故障模式将变得司空见惯。该研究计划的主要目标是设计并行算法和多物理CFD解算器，以支持在亿级星系团上进行下一代CFD模拟。我将重点介绍光滑粒子流体动力学(SPH)，这是一种广泛使用的无网格CFD方法，以及它在天体物理中的应用。我将集中在两个主要领域：为艾级星系团设计并行算法，以及通过艾级计算创建解决关键天体物理问题所需的多物理算法。第一个艾级星系团将很快成为现实，因为美国国家战略计算倡议的目标是到2023年拥有一个有能力的艾级星系团。这项研究计划的结果将对加拿大社会产生许多有益的影响，因为利用亿级资源的稳健算法将在许多领域有用，如风力发电场优化、抗癌药物发现和核反应堆设计。通过我的研究计划对本科生和研究生进行培训，将培养出具有宝贵技能的高素质人才(HQP)，如软件工程和数据科学，这些技能可以转移到广泛的职业领域。

Fluids are everywhere in nature. Computational fluid dynamics (CFD) is the broad field of using computers to obtain solutions to the equations describing fluid flow. Performing simulations of fluid flow using CFD techniques has become an indispensable tool to study problems across a range of scientific disciplines. Simulations allow us to validate of our understanding of complex phemonena, and, in many cases, offers the only potential avenue to study problems that are intractable from a mathematical or experimental perspective. Simulations continually increase in scope, reflecting the need to connect small-scale processes with large-scale features, and the drive to create wholly consistent simulations that encompass all relevant physical processes. That simulations can increase in scale and complexity is permissible because supercomputing clusters continue to exponentially grow in computational power over time. Within the next decade, the first supercomputing clusters will come online that can compute at 10^18 floating point operations per second (an exaflop). However, there exist substantial challenges in designing parallel algorithms that can scale CFD solvers for exascale clusters. These clusters will have more computational power not by individual CPUs becoming faster, but by sheer number of CPUs, and by including a mixture of CPUs and GPUs (graphics processing units). Communication loads and memory bottlenecks will need to be overcome, and failure modes that are rare enough to ignore today will become commonplace. The primary objectives of this research program are to design parallel algorithms and multi-physics CFD solvers that enable the next generation of CFD simulations on exascale clusters. I will focus on smoothed particle hydrodynamics (SPH), a widely used mesh-free CFD method, and its applications to astrophysics. I will focus on two main areas: designing parallel algorithms for exascale clusters, and creating the multi-physics algorithms required to solve key astrophysical problems by exascale computation. The first exascale clusters will soon be reality, as the US National Strategic Computing Initiative aims to have a capable exascale cluster by 2023. The outcomes of this research program will have numerous beneficial impacts to Canadian society, as robust algorithms to utilize exascale resources will be useful in many areas, such as wind farm optimization, cancer drug discovery, and nuclear reactor design. Training of undergraduate and graduate students through my research program will yield highly-qualified personnel (HQP) with valuable skills, such as software engineering and data science, that are transferrable to a broad range of careers.