SPX: Collaborative Research: Cross-layer Application-Aware Resilience at Extreme Scale (CAARES)

科学和工程应用日益增长的需求推动了当前大规模系统的极限，并有望在未来十年内实现百亿亿次（10^18 FLOPS）的性能。在极端规模下，研究较少的挑战之一是计算系统本身的可靠性，这主要是由于所使用的内核和组件数量非常大，并且这些系统上的平均故障间隔时间急剧减少（大约几十分钟）。该项目脱离了传统的单组件故障管理模型，并探索了在单个并行应用程序上下文中使用的多个软件库（和应用程序组件）如何相互作用，以提供针对能力计算的并行应用程序所需的整体故障管理支持。这种探索将不局限于使用单一并行编程范式开发的软件，而是将扩展到包含更具挑战性的情况，即多种编程范式可以结合起来实现共同目标，以模拟当今使用的一组大规模科学应用程序。该项目的目标是摆脱目前孤立的弹性机制，并提出跨层组合解决方案，可以从根本上解决极端规模下的这些弹性挑战。这种探索将不局限于使用单一并行编程范式开发的软件，而是将扩展到包含更具挑战性的情况，即多种编程范式可以结合起来实现共同目标，以模拟当今使用的一组大规模科学应用程序。更具体地说，本提案将解决以下研究挑战：(1)为更深入地理解结合不同弹性模型和方法所带来的挑战和机遇奠定理论基础；(2)设计一个灵活的编程抽象，允许不同的弹性模型和机制相结合，以更全面的方式合作和解决弹性问题；(3)开发基本的、独立于编程范式的、必要的结构，以实现跨层和特定于领域的方法，以支持弹性和理解相关的性能/质量权衡。通过在两种不同的编程范式（MPI和OpenSHMEM）中公开这些通用抽象，并为每种范式创建和开发专门的概念，可以验证所建议的方法。这将允许使用很少的软件框架和应用程序来评估概念的有效性和由不同软件层施加的相应开销。

The increasing demands of science and engineering applications push the limits of current large-scale systems, and is expected to achieve exascale (10^18 FLOPS) performance early in the next decade. One of the lesser studied challenge at extreme scales is the reliability of the computing system itself, primarily due to the very large number of cores and components utilized and to the sharp decrease of the Mean Time Between Failures on such systems (in the order of tens of minutes). This project departs from the traditional single component fault management model, and explores how multiple software libraries (and application components) used in the context of a single parallel application can interact to provide the holistic fault management support necessary for parallel applications targeting capability computing. This exploration will not be limited to software developed using a single parallel programming paradigm, but will be extended to encompass the more challenging case where multiple programming paradigms can be combined to achieve a common goal, to simulate a set of large scale scientific applications in use today.  The goal of this project is to depart from the current siloed resilience mechanisms, and propose cross-layer composition solutions that can fundamentally address these resilience challenges at extreme scales. This exploration will not be limited to software developed using a single parallel programming paradigm, but will be extended to encompass the more challenging case where multiple programming paradigms can be combined to achieve a common goal, to simulate a set of large scale scientific applications in use today. More specifically, this proposal will address the following research challenges: (1) development of a theoretical foundation for a deeper understanding of the challenges and opportunities arising from combining different resilience models and methodologies; (2) design of a flexible programming abstraction to allow different resilience models and mechanisms to be combined to cooperate and address resilience in a more holistic manner; and (3) development of basic, programming paradigm independent, constructs necessary to implement cross-layer and domain-specific approaches to support resilience and to understand related performance / quality trade-offs. The proposed approach will be validated by exposing these generic abstractions in two different programming paradigms (MPI and OpenSHMEM), by creating and developing specialized concepts for each of these paradigms. This will enable the assessment of the validity of the concepts and the corresponding overheads imposed by the different software layers, using few software frameworks and applications.