CAREER: Cross-Layer Power-Bounded High Performance Computing on Emerging and Future Heterogeneous Computer Clusters

高效、可扩展的计算系统对科学发现和技术创新至关重要，对国家安全和人类社会至关重要。然而，高性能计算系统的可扩展性越来越受到功率要求和限制组件和服务器室功率密度的必要性的限制。由数以百万计的部件组成？HPC系统已经消耗了兆瓦级的电力；为了满足关键任务应用程序对性能的永不满足的需求，未来的系统将由更多的组件组成，并消耗更多的功率。为了解决扩展性能和限制功率的冲突需求，本研究开发了在新兴和未来受功率预算限制的计算机系统上实现高效和可扩展计算的使能技术。提出的功率限制HPC方法将功率视为稀缺资源，并利用硬件过度配置在功率预算内扩展性能。针对新兴的异构HPC集群，包括功率感知的多核cpu和多核加速器，本研究研究如何利用所有可用的功率来最大化组件，节点和集群级别的性能和功率效率，用于广泛的应用。具体而言，本研究(1)设计了一种新型的应用感知跨组件调度系统，用于功率有界多核计算；(2)创建了一种用于功率有界异构计算的协同混合计算框架；(3)开发了支持大规模功率有界计算的分析模型和技术。本研究的完成促进了新颖的系统和软件设计，有效地利用每瓦特的计算功率；由此产生的分析模型为设计未来的高性能计算系统、体系结构和构建模块奠定了理论基础。该项目整合了教育内容，吸引研究生和本科生参与创新的高性能计算研究，并扩大代表性不足和K-12学生的参与。

Highly efficient and scalable computing systems are crucial to scientific discovery and technology innovation critical to national security and human society. However, the scalability of HPC systems is increasingly constrained by the power requirement and the necessity to limit the power density of components and server rooms. Comprising millions of components, today?s HPC systems already consume megawatts of power; to meet an insatiate demand for performance from mission-critical applications, future systems will consist of even more components and consume more power. To resolve the conflicting needs of scaling performance and limiting power, this research develops enabling technology for efficient and scalable computing on emerging and future computer systems bounded by power budgets.The proposed power-bounded HPC approach recognizes power as a scarce resource and exploits hardware overprovisioning to scale performance within a power budget. Targeting at emerging heterogeneous HPC clusters comprising power-aware multicore CPUs and manycore accelerators, this research studies how to utilize all available power to maximize performance and power efficiency at component, node, and cluster levels for a wide range of applications. Specifically, this research (1) designs a novel application-aware cross-component scheduling system for power-bounded multicore computing, (2) creates a cooperative hybrid computing framework for power-bounded heterogeneous computing, and (3) develops analytical models and techniques to support large scale power-bounded computing. The completion of this research promotes novel system and software designs that efficiently utilize every watt of power on computation; the resulting analytical models form the theoretical foundation for designing future HPC systems, architectures, and building blocks. This project integrates educational components that engage graduate and undergraduate students in innovative HPC research, and broaden the participation of underrepresented and K-12 students.