Collaborative Research: Performance Toolset for Dynamic Optimization of High-End Hybrid Applications

目前的高端应用程序通常只利用大型平台理论性能的一小部分。硬件、系统软件、编程接口和算法之间的交互极其复杂，对于混合MPI+OpenMP Fortran/C/C++应用程序的开发具有挑战性。新一代的机器将更加复杂，利用它们的应用程序也将更加复杂。典型的应用程序开发和调优场景涉及手动和单独使用编译器和性能工具，以及根据从其输出中辛苦收集的见解修改程序。在这项提案中，我们打算通过创建一个集成的程序优化环境来提高应用程序开发和调优过程的质量，该环境可以减少现有方法的人工劳动和猜测。我们将开发策略和相应的接口，使应用程序开发人员、编译器和性能工具能够协作，根据各种反馈来源(包括来自.Offline的性能数据)生成优化的代码。开发可以从.Online运行，也可以从.Online运行。生产运转。我们将构建和部署一个灵活的、工作的系统，将强大的、现有的、开放源码的软件结合在一起。将一个编译器、一个程序分析工具和两个具有互补功能的性能工具整合到一个统一的环境中，以进行协作的静态和动态应用程序调优。由我们的应用程序合作伙伴提供的各种复杂程度的应用程序代码将激励我们的开发工作，并测试和展示我们的结果。其结果将是一个强大的集成环境，可用于通过程序监视、事件跟踪和/或硬件计数器信息的提取来获得传统的性能数据，并获得对应用程序代码的静态或动态调整的支持。智能优点建议的环境集成了几种不同的现有技术，为优化混合MPI+OpenMP代码提供了一个新的支持水平。提供了对混合编程模型实验的支持。编译器可以用来优化代码的信息范围被扩展到涵盖许多系统和应用程序级别的现象以及各种优化方案。工具之间的交互将有助于提供一种能够处理极端规模计算的方法。将解决集成问题，目标是创建一个可部署、可扩展的系统。更广泛的影响本项目的工具、想法和成果将免费分发，并在国内和国际上向HPC社区提供。除了成果的广泛传播，该项目还与NCSA的性能工程专家有很强的联系。这些工具将安装在NCSA系统上进行测试和评估，并将向其他用户提供。这项研究汇集了编译器、性能工具和应用程序开发人员，丰富了研究生的研究经验，创造了一支全面发展的IT劳动力队伍。PI活跃在OpenMP社区中，项目团队与美国能源部和美国国防部有着密切的工作关系。产生的新知识将被整合到高级研究生课程中。

Current high-end applications usually exploit just a fraction of the theoretical performance of large platforms. Interactions among the hardware, system software, programming interface, and algorithm are extremely complex, and the implications for development of hybrid MPI+OpenMP Fortran/C/C++ applications are challenging. The emerging generation of machines will be even more complex, as will the applications that exploit them. Typical application development and tuning scenarios involve the manual and separate use of compilers and performance tools, and program modifications based upon insights laboriously gleaned from their output. In this proposal, we intend to raise the quality of the application development and tuning process by creating an integrated environment for program optimization that reduces the manual labor and guesswork of existing approaches. We will develop strategies and the corresponding interfaces that enable the application developer, compiler and performance tools to collaborate to generate optimized code based upon a variety of sources of feedback, including performance data from .offline. development runs as well as from .online. production runs. We will build and deploy a flexible, working system that combines robust, existing, open source software . a compiler, a program analysis tool and two performance tools with complementary features - into a single, coherent environment for collaborative static and dynamic application tuning. Application codes of varying complexity supplied by our application partner will motivate our development work as well as test and demonstrate our results. The result will be a powerful, integrated environment that can be used to obtain traditional performance data via program monitoring, event tracing and/or the extraction of hardware counter information, and to obtain support for the static or dynamic tuning of an application code. Intellectual Merit The proposed environment integrates several different existing technologies to provide a new level of support for optimizing hybrid MPI+OpenMP codes. Support for experimentation with the hybrid programming model is provided. The range of information that may be exploited by the compiler to optimize code is expanded to cover many system and application-level phenomena and a variety of optimization scenarios. Interactions between tools will facilitate the provision of an approach that is able to handle extreme-scale computations. Integration issues will be addressed with the goal of creating a deployable, extensible system. Broader Impacts The tools, ideas, and results of this project will be freely distributed and made available to the HPC community, nationally and internationally. Besides the general dissemination of results, the project has strong ties to performance engineering experts at NCSA. The tools will be installed on NCSA systems for testing and evaluation and will be made available to other users. The research brings together compiler, performance tool, and application developers, enriching the research experience of graduate students to create a well-rounded IT workforce. The PI is active in the OpenMP community and the project team has close working relationships with DOE and DOD. New knowledge generated will be integrated into advanced graduate coursework.