CDS&E: Optimization of Advanced Cyberinfrastructure through Data-Driven Computational Modeling

现代科学和大数据计算系统必须平衡功耗、性能、可靠性等多个系统属性，以满足应用科学的需求。然而，这些系统及其设计者受到缺乏清晰和定义良好的方法来指导系统设计和调优的限制。这降低了现代战略计算系统的整体效率。该项目正在开发新技术，用于定量表征和优化广泛的现代战略计算系统中的系统权衡。这个项目的短期目标是开发模型、分析和优化，使系统软件、应用程序和系统架构师能够做出有效的端到端性能权衡。本研究的预期长期影响是提高当前和新兴战略计算系统的整体效率和有效性。作为该项目的一部分开发的技术也将与新墨西哥大学和全国各地的大规模计算机教育项目相结合。本研究以描述应用程序/系统-软件配置性能的随机集体-区间模型为基础。为此，该项目正在研究评估不同系统软件和应用程序机制如何改变应用程序模型参数的技术，为理解应用程序/系统软件性能权衡提供一种通用方法。此外，该项目正在研究利用这些模型来改进系统调度器和运行时系统中的资源分配决策的优化技术。评估这些技术的依据是它们在调度、电源管理和数据局部性等领域改善现代战略计算系统整体性能的能力。作为该项目的一部分开发的技术也将与新墨西哥大学和全国各地的大规模计算机教育项目相结合。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。

Modern scientific and big-data computing systems must balance multiple system attributes such as power, performance, and reliability to meet application science demands. These systems and their designers are, however,constrained by the lack of clear and well-defined methods to guide system design and tuning. This reduces the overall effectiveness of modern strategic computing systems. This project is developing newtechniques for quantitatively characterizing and optimizing system tradeoffs in a wide range ofmodern strategic computing systems. The short term goal of this project is to develop models, analyses,and optimizations that will enable system software, applications, and system architects to make effective end-to-end performance tradeoffs. The desired long-term impact of this research is to increase theoverall efficiency and effectiveness of current and emerging strategic computing systems. The techniquesdeveloped as part of this project will also be integrated with large-scale computing educational programsat the University of New Mexico and across the country.This research is grounded in stochastic inter-collective-interval model for characterizing the performance of application/system-software configurations. To this end, the project is investigating techniques for estimating how different system software and applications mechanisms change application model parameters, providing a general means for understanding application/system software performance tradeoffs. In addition, the project is examining optimization techniques that leverage these models to improve resource allocation decisions in system schedulers and runtime systems. These techniques are being evaluated based on their ability to improve the overall performance of modern strategic computing systems in areas such as scheduling, power management, and data locality. The techniquesdeveloped as part of this project will also be integrated with large-scale computing educational programsat the University of New Mexico and across the country.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.