Collaborative Research: PPoSS: Planning: SEEr: A Scalable, Energy Efficient HPC Environment for AI-Enabled Science

人工智能使能的科学，其中先进的机器学习技术被用于代理模型、自动调整和现场数据分析，正迅速在科学和工程中被采用，以解决复杂和具有挑战性的计算问题。嵌入不同类型处理设备(CPU、GPU和AI加速器)的异类系统的广泛采用使在超级计算机上执行支持AI的科学变得更加复杂。在异质系统上实现人工智能仿真的研究还远远不够。该项目的新颖性是探索对于在不同系统上实现人工智能科学的可扩展、高能效的HPC环境所必需的关键功能。统一的研究团队以跨层的方式解决这个问题，重点关注应用程序算法、编程语言和编译器、运行时系统和高性能计算之间的协同效应。该项目的影响是通过使科学计算更快、更具可扩展性和更节能来催化科学发现。长期研究愿景是开发SEER，这是一个可扩展的、高能效的HPC环境，用于扩大和加速支持人工智能的科学，以促进科学发现。这个规划项目探索了实现研究愿景的基本问题。该团队专注于使用OpenFOAM的不可压缩计算流体动力学应用程序的可扩展代理模型、该应用程序在不同资源上的成本模型、用于有效执行的动态任务映射以及性能和功率监控与表征，以探索性能、可扩展性和能效之间的权衡，并在名为Polaris的最先进的试验台上进行。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

AI-enabled science, where advanced machine-learning technologies are used for surrogate models, autotuning, and in situ data analysis, is quickly being adopted in science and engineering for tackling complex and challenging computational problems. The wide adoption of heterogeneous systems embedded with different types of processing devices (CPUs, GPUs, and AI accelerators) further complicates the execution of AI-enabled science on supercomputers. The research for AI-enabled simulations on heterogeneous systems is far from sufficient. The project’s novelty is to explore key features essential for a scalable, energy-efficient HPC environment for AI-enabled science on heterogeneous systems. The unified team of researchers tackles the problem in a cross-layer manner, focusing on the synergies among application algorithms, programming languages and compilers, runtime systems, and high-performance computing. The project's impact is to catalyze scientific discoveries by making scientific computing faster, more scalable and more energy-efficient. The long-term research vision is to develop SEEr, a scalable, energy-efficient HPC environment for scaling up and accelerating AI-enabled science for scientific discovery. This planning project explores fundamental questions to realize the research vision. The team focuses on scalable surrogate models for an incompressible computational fluid dynamics application using OpenFOAM, cost models for this application on heterogeneous resources, dynamic task mapping for efficient execution, and performance and power monitoring and characterization to explore tradeoffs among performance, scalability, and energy efficiency on a state-of-the-art testbed named Polaris.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.