CSR :Small: Exploiting Slowdowns for Speedup in Power-Scalable HPC Systems.

高级计算系统——支持经济、科学和医学等领域的广泛应用——在设计时越来越多地考虑到能源效率。现有的能源管理方法是在不同的电压和频率下运行底层处理器和设备。通常情况下，这些方法会在“越快越好，或者至少不会造成伤害”的前提下，推动设备在热限制下尽可能快地运行。在主流文献中，越来越多的证据表明“有时越慢越好”。例如，对于IOZone这样的基准测试应用程序，已经观察到以更快的速度运行处理器可能会导致总体执行时间的显着减慢。在大规模的情况下，例如在Amazon Web Services云中，这样的性能损失可能会花费数十万美元的CPU小时，并浪费宝贵的能源，这些能源通常来自污染性的化石燃料。然而，在当今数据中心规模的复杂系统中，找出这种减速的根本原因就像是大海捞针。性能现在是应用程序设计、系统资源和底层硬件之间复杂交互的一个功能。此外，功率缩放使硬件的原始性能成为一个变量；因此，进一步混淆了隔离减速的尝试。该项目构建了新的技术，在启用功率扩展时，可以识别、建模和自动化最小化或消除并行和分布式应用程序中的减速。关键方法是细粒度应用程序和内核检测，以深入分析并行和分布式应用程序与软件和硬件堆栈之间的交互。本研究的智力价值涉及三个中间研究目标：1)对大量应用程序和各种系统进行详尽的测试和深入的系统和代码分析，以分类和隔离由于功率缩放导致的减速现象；2)设计、实现和验证对减速敏感的应用程序关键路径模型；3)分析生成的模型以及自动化、开源、运行时优化技术的设计、实现和验证，以引导功率缩放以最小化或消除减速。该项目的完成将提高先进系统的性能和能源效率。采用由此产生的运行时工具将支持使用功率缩放来节省能源，同时减少现代应用程序和系统的解决方案时间。由此产生的人工制品和技术将通过解决在电力限制下建造大型系统的挑战，为美国的竞争力做出贡献。这些教育活动将有助于培养具有高度市场需求技能的多元化毕业生。将研究发现和软件工具整合到教育课程中，这将是开源和公开的，将有助于吸引下一代计算机科学家的兴趣。

Advanced computing systems --- that support a wide variety of applications in fields such as economics, sciences, and medicine --- are increasingly being designed with energy efficiency considerations. An extant approach to energy management is to run the underlying processors and devices at varying voltage and frequency. Typically, the approaches push the devices to run as fast as possible within thermal limits using the premise that "faster is better or at least does no harm." There is growing evidence in the prevailing literature that "slower is sometimes better." For example, for benchmark applications such as IOZone, it has been observed that running the processors at a faster speed can lead to significant slowdowns in the overall execution time. At large scale, e.g., in the Amazon Web Services cloud, such performance loss can cost hundreds of thousands of dollars in CPU hours and waste precious energy often begotten from polluting fossil fuels. However, isolating the root cause of such slowdowns in today's complex systems at the scale of data centers is akin to finding a needle in a haystack. Performance is now a function of the complex interaction between application design, system resources, and the underlying hardware. Furthermore, power scaling makes the raw performance of the hardware a variable; thus, further confounding attempts to isolate slowdowns.This project builds novel technologies that identify, model and automate the minimization or elimination of slowdowns in parallel and distributed applications when power scaling is enabled. The key approach is fine-grain application and kernel instrumentation to develop in-depth analysis of the interaction between parallel and distributed applications and the software and hardware stack. The intellectual merit of this research involves three intermediate research goals: 1) Exhaustive testing and deep system and code analysis on a large class of applications and a diverse set of systems to classify and isolate the slowdown phenomenon due to power scaling; 2) Design, implementation, and validation of models of the critical paths of applications exhibiting sensitivity to slowdowns; and 3) Analysis of the resulting models and design, implementation, and validation of the automated, open-source, runtime optimization techniques to steer power scaling to minimize or eliminate slowdowns.Completion of the project will improve the performance and energy efficiency of advanced systems. Adoption of the resulting runtime tools will enable use of power scaling to save energy while simultaneously reducing time-to-solution for modern applications and systems. The resulting artifacts and technologies will contribute to U.S. competitiveness by addressing the challenge of building large-scale systems within power constraints. The educational activities will help produce diverse graduates with highly marketable skill sets. The integration of the research discoveries and software tools, which will be open source and made public, into the educational curriculum will help capture the interest of the next generation of computer scientists.