SI2-SSI: Collaborative Proposal: Performance Application Programming Interface for Extreme-Scale Environments (PAPI-EX)

现代高性能计算(HPC)系统在尺寸和复杂性上持续增加。在这些日益复杂的环境中衡量应用程序性能的工具还必须增加其测量的丰富性，以深入了解软件和硬件交互的日益复杂的方式。PAPI性能监控库为所有现代CPU和一些其他相关组件(散布在芯片和系统中)上可用的硬件性能计数器提供了一个清晰、可移植的接口。被广泛部署和广泛使用的PAPI已经成为基本的软件基础设施，不仅适用于科学计算社区，也适用于许多高性能计算的行业用户。但过去几年在处理器和系统设计方面发生的根本性变化对PAPI和整个HPC软件基础设施提出了新的挑战。PAPI-EX项目集成了来自政府和行业研究投资的关键PAPI增强，重点关注处理器和系统设计更改，这些更改预计将出现在通往亿级计算的每个极端规模平台上。PAPI-EX的主要影响直接取决于PAPI库的重要性。15年来，工具开发人员、主要的国家HPC中心、系统供应商和应用程序开发人员一直主要使用PAPI。Papi-ex建立在这个基础上。作为重要的研究基础设施，PAPI-EX项目使PAPI能够在新系统的设计和规模发生革命性变化的情况下继续发挥其重要作用。在加强发现和教育方面，与PAPI-EX合作的合作伙伴名单包括NSF计算中心、主要工具开发商、主要系统供应商和个人社区领导人，这一多样化的群体将有助于促进培训课程、有针对性的研讨会以及在国家和国际会议上的小型专题讨论会。最后，许多主要系统供应商积极推广PAPI，这意味着PAPI-EX将继续在许多领域为政府和行业带来重大好处。PAPI-EX针对的是一种硬件环境，在该环境中，当前和未来的多核CPU的核心共享各种性能关键资源(也称为“核间”资源)，包括电源管理、片上网络、存储器层次结构和核心之间的存储器控制器。未能管理对这些“核心间”资源的争用已经成为拖累整体应用程序性能的主要因素。因此，对于许多性能工具(例如，TAU、HPCToolkit、Open|Speedshop、Vampir、Scalasca、CrayPat、Active Harmony等)的用户来说，缺乏在低级别揭示这些资源的实际行为的能力已经变得非常成问题。PAPI-EX增强和扩展了PAPI以解决这一关键问题，并为其在高性能计算性能优化中发挥其公认的作用做好准备。因此，PAPI-EX的目标如下：(1)开发包括系统范围和核心间测量的共享硬件计数器支持；(2)为基于数据流的运行时系统提供支持；(3)创建采样接口以记录具有相关上下文的性能数据流；(4)将易于使用的基于文本的应用程序性能分析工具与PAPI？S高级API的更新相结合，以创建一个基本的、开箱即用的？检测API。

Modern High Performance Computing (HPC) systems continue to increase in size and complexity. Tools to measure application performance in these increasingly complex environments must also increase the richness of their measurements to provide insights into the increasingly intricate ways in which software and hardware interact. The PAPI performance-monitoring library has provided a clear, portable interface to the hardware performance counters available on all modern CPUs and some other components of interest (scattered across the chip and system). Widely deployed and widely used, PAPI has established itself as fundamental software infrastructure, not only for the scientific computing community, but for many industry users of HPC as well. But the radical changes in processor and system design that have occurred over the past several years pose new challenges to PAPI and the HPC software infrastructure as a whole. The PAPI-EX project integrates critical PAPI enhancements that flow from both governmental and industry research investments, focusing on processor and system design changes that are expected to be present in every extreme scale platform on the path to exascale computing.The primary impact of PAPI-EX is a direct function of the importance of the PAPI library. PAPI has been in predominant use by tool developers, major national HPC centers, system vendors, and application developers for over 15 years. PAPI-EX builds on that foundation. As important research infrastructure, the PAPI-EX project allows PAPI to continue to play its essential role in the face of the revolutionary changes in the design and scale of new systems. In terms of enhancing discovery and education, the list of partners working with PAPI-EX includes NSF computing centers, major tool developers, major system vendors, and individual community leaders, and this diverse group will help facilitate training sessions, targeted workshops, and mini-symposia at national and international meetings. Finally, the active promotion of PAPI by many major system vendors means that PAPI, and therefore PAPI-EX, will continue to deliver major benefits for government and industry in many domains.PAPI-EX addresses a hardware environment in which the cores of current and future multicore CPUs share various performance-critical resources (a.k.a., 'inter-core' resources), including power management, on-chip networks, the memory hierarchy, and memory controllers between cores. Failure to manage contention for these 'inter-core' resources has already become a major drag on overall application performance. Consequently, the lack of ability to reveal the actual behavior of these resources at a low level, has become very problematic for the users of the many performance tools (e.g., TAU, HPCToolkit, Open|SpeedShop, Vampir, Scalasca, CrayPat, Active Harmony, etc.). PAPI-EX enhances and extends PAPI to solve this critical problem and prepare it to play its well-established role in HPC performance optimization. Accordingly, PAPI-EX targets the following objectives: (1) Develop shared hardware counter support that includes system-wide and inter-core measurements; (2) Provide support for data-flow based runtime systems; (3) Create a sampling interface to record streams of performance data with relevant context; (4) Combine an easy-to-use tool for text-based application performance analysis with updates to PAPI?s high-level API to create a basic, ?out of the box? instrumentation API.