SHF:Small:Intelligent Management of Hybrid Workloads for Extreme Scale Computing

高性能计算（HPC）社区正在将人工智能（AI）技术用于无数的追求，从推动突破性的科学发现到保护我们的国家安全。随着新出现的机器学习和以数据为中心的工作负载在HPC中激增，当前的工作负载管理系统无法跟上在异构系统上共同运行的各种应用程序组合所带来的重大挑战。该项目通过开发新的工作负载管理方法来解决这个问题，从而促进高性能计算、人工智能和数据分析的融合。它将在高性能计算工作负载管理方面进行根本性改进，以促进大规模超级计算机在新兴数据中心应用（HPC4AI）中的使用。同时，它将利用先进的人工智能技术，特别是多目标强化学习，为高性能计算（AI4HPC）中的作业调度和资源分配提供支持。该项目旨在开发一个名为MINT的智能工作负载管理框架，在该框架中，混合工作负载的不同计算资源需求将被自动识别和满足，以实现极高的资源效率和满意的用户体验。主要研究重点是：通过模型驱动分析了解超级计算机上不同工作负载的性能影响；智能多资源调度新方法最小化工作负载干扰的智能资源分配策略并通过基于跟踪的仿真和测试对所提出的框架进行了广泛的评估。交付成果包括一个新的工作负载管理框架和开源软件版本，用于在极端规模系统上智能管理混合工作负载。该奖项反映了美国国家科学基金会的法定使命，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。

The high-performance computing (HPC) community is embracing artificial intelligence (AI) techniques for countless pursuits, from driving ground-breaking scientific discoveries to protecting our national security. As newly emerging machine learning and date-centric workloads proliferate in HPC, current workload-management systems cannot keep up with the significant challenges introduced by the diverse mix of applications co-running on heterogeneous systems. This project tackles the problem by developing new workload-management methods to catalyze the convergence of HPC, AI, and data analytics. It will develop fundamental improvements in HPC workload management to promote the use of large-scale supercomputers for emerging data-centric applications (HPC4AI). Meanwhile it will exploit advanced AI technologies, especially multi-objective reinforcement learning, to empower job scheduling and resource allocation in HPC (AI4HPC). The project aims to develop an intelligent workload-management framework named MINT in which distinctive computational resource requirements of hybrid workloads will be automatically identified and fulfilled to achieve extreme resource efficiency and satisfactory user experience. Key research thrusts are: understanding performance implications of diverse workloads on supercomputers via model-driven analysis; new intelligent multi-resource scheduling methods; smart resource-allocation strategies for minimal workload interference; and extensive evaluation of the proposed framework through trace-based simulation and testing. The deliverables include a new workload-management framework and open-source software releases for intelligent management of hybrid workloads on extreme-scale systems.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.