CSR: Rethinking System Software for Overprovisioned, High-Performance Computing Systems

目前，高性能计算（HPC）社区专注于实现exaflop的性能，这比当今世界上最好的超级计算机的性能提高了30倍。出于实际、财政和环境方面的考虑，美国能源部正在为达到20兆瓦的百亿亿次浮点运算设定功率限制。由于当今顶级计算机的功耗通常在5到20兆瓦之间，而且距离exaflop的性能目标还有一个数量级或更多，因此HPC系统的重大硬件和软件进步是必要的。改进硬件的一种方法是使用过度配置的系统，其中包含的机器数量超过了同时完全供电的能力。虽然过度配置的系统具有显著提高功率和性能的潜力，但需要重新设计软件以支持此类系统。这个建议的重点是设计和实现支持过度供应系统的软件基础设施。基础设施的关键进步是支持系统范围的优化，即跨多个应用程序的优化。这与当前HPC系统以每个应用程序为基础进行优化的重点形成鲜明对比。开发的软件将包括作业分析器、一次对多个作业执行分析的调度器和基于调度器分析输出联合优化多个应用程序的集群范围运行时系统。实现百亿亿次计算是国家的重要优先事项，将影响许多关键应用领域，如气候/天气、可再生能源、核能、材料科学和国家安全。本文描述的工作将提高功率受限的高性能计算系统的整体性能，这是迈向百亿亿级目标的重要一步。该项目计划通过与几个国家实验室的长期合作，以拟议的软件堆栈的形式转让这项研究产生的技术。

Currently, the high-performance computing (HPC) community is focused on achieving exaflop performance, which is about a 30-fold improvement from the performance of the best supercomputer in the world today. Because of practical, financial, and environmental concerns, the Department of Energy is setting a power limit for achieving an exaflop at 20 megawatts. As today's top machines generally consume between five and 20 megawatts---and yet are an order of magnitude or more away from the exaflop performance target, significant hardware and software advances in HPC systems are necessary. One way to improve hardware is to use overprovisioned systems, which contain more machines than can be fully powered simultaneously. While overprovisioned systems have the potential to significantly improve power and performance, software will need to be redesigned to support such systems.The focus of this proposal is to design and implement software infrastructure that will support overprovisioned systems. The key advance in the infrastructure is support of system-wide optimizations, i.e., optimizations that span multiple applications. This is in stark contrast to the current focus in HPC systems of optimizing on a per-application basis. The developed software will consist of a job profiler, a scheduler that performs analysis on multiple jobs at a time, and a cluster-wide run-time system that jointly optimizes multiple applications based on the output of the scheduler analysis.Achieving exascale computing is an important national priority and will impact many critical application domains, such as climate/weather, renewable energy, nuclear energy, materials science, and national security. The work described here will improve whole-system performance on power-constrained HPC systems, which is one important step towards the exascale goal. The project plans to transfer technology resulting from this research in the form of the proposed software stack via longstanding collaborations with several national laboratories.