Collaborative Research: ITR/NGS: Deja Vu: Transparent Checkpointing and Migration of Parallel Codes Over Grid Infrastructures

一个艰巨的挑战是从今天的计算网格演变为真正的网络基础设施，它无缝地整合了从学术实验室的小集群到最大的国家超级计算中心的各种资源，并提供对高性能计算、研究仪器、数据仓库和可视化的无处不在的访问。要实现这一未来，需要在透明故障恢复机制方面取得根本性进展，以屏蔽任何大规模计算资源特有的组件故障。前几代超级计算机在系统硬件中设计了可靠性，而今天的高性能计算(HPC)环境是基于COTS组件的集群，没有针对资源整体可靠性的系统解决方案。为了在不断增长的集群系统网络集合中产生稳定性，需要一种软件解决方案，该解决方案通过透明、高效和自动检查点设置和恢复(CPR)机制提供对计算资源的可靠访问。该项目旨在通过对CPR中长期存在的问题采取全新的方法实现这一未来，并通过建立一个称为似曾相识的综合系统来实现流程迁移。似曾相识提供(A)透明的并行检查点和恢复机制，可从任何系统故障组合中恢复，而无需对并行应用程序进行任何修改。(B)一种透明地捕获应用程序状态的新型编译器后分析系统，(C)在单个框架中无缝集成用户发起的检查点和系统发起的检查点的系统体系结构，使得能够有效地使用各种特定领域的知识，(D)用于透明增量检查点的新型运行时机制，以有效地捕获维护全局一致性所需的最少量的状态，(E)新型通信体系结构，其使得能够透明地将现有MPI/PVM代码迁移到应用程序或MPI/PVM库，(F)可针对特定存储环境量身定制的可恢复IO子系统，和(G)Globus工具包的接口和扩充，以有效利用这项研究提供的CPR和迁移能力。Deja vu的核心CPR和迁移设施将被管理、安全和调度设施包围，这些设施(A)与本地调度系统(例如OpenPBS)和会计系统相结合，以便对丢失的计算周期进行特定地点的核算和退款，以及(B)扩展Globus安全架构，提供细粒度权利和动态创建的用户账户，使Deja vu系统下可用的流体资源控制得到充分利用。该项目的设计目标不仅是实现“点”解决方案，而且是一个集成系统，它将构成大型计算设施和网格基础设施的基本组件。我们的研究团队(VT、PSC、ISR)在完整解决方案的设计、开发、部署和支持方面拥有丰富的经验。

A daunting challenge is the evolution from today's computational Grid to a true cyberinfrastructure that seamlessly integrates resources ranging from small clusters in academic laboratories to the largest national supercomputing centers and provides ubiquitous access to high performance computing, research instrumentation, data warehouses and visualization. Realization of this future requires fundamental advances in transparent fault recovery mechanisms to mask component failures endemic to any large-scale computational resource. While previous generations of supercomputers engineered reliability into systems hardware, today's high performance computing (HPC) environments are based on clusters of COTS components, with no systemic solution for the reliability of the resource as a whole. Engendering stability in ever growing networked collections of cluster systems needs a software solution that provides reliable access to computing resources through transparent, efficient, and automatic checkpointing and recovery (CPR) mechanisms. This project aims to bring about this future through radically new approaches to longstanding problems in CPR and process migration by building an integrated system called Deja vu. Deja vu provides (a) a transparent parallel checkpointing and recovery mechanism that recovers from any combination of systems failures without any modification to parallel applications. (b) a novel post-compiler analysis system that transparently captures application state, (c) a systems architecture that seamlessly integrates user-initiated and system-initiated checkpoints in a single framework enabling the effective use of a wide variety of domain specific knowledge, (d) novel runtime mechanisms for transparent incremental checkpointing, to efficiently capture the least amount of state required to maintain global consistency, (e) a novel communications architecture that enables transparent migration of existing MPI/PVM codes without source-code modifications to either the application or the MPI/PVM libraries, (f) recoverable IO subsystems that can be tailored to specific storage environments, and (g) interfaces to and augmentation of the Globus Toolkit to effectively use the CPR and migration capabilities provided by this research. The core CPR and migration facilities of Deja vu will be surrounded by management, security, and scheduling facilities that (a) integrate with local scheduling systems (e.g., OpenPBS) and accounting systems for site-specific accounting and refunding of lost compute cycles and (b) extend the Globus security architecture with fine grain rights and dynamically created user accounts that allow the fluid resource control available under the Deja vu system to be fully exploited. The design goal of this project is not just to implement "point" solutions, but an integrated system that will constitute a fundamental component of both large-scale computing facilities and Grid infrastructures. Our research team (VT, PSC, ISR) has considerable experience in the design, development, deployment and support of complete solutions.