Design and Evaluation of System-Level Fault Tolerance Mechanisms for Parallel Computers

最近的发展极大地增加了超可靠并行计算的重要性。首先，并行处理已经从专门的科学计算工具转变为通用计算工具。与此同时，计算机已经开始控制人类生活的几乎每一个方面，因此，人们越来越依赖计算机来提供持续的服务。这些趋势使得开发通用但实用的容错多处理技术成为必要。这个项目的重点是松耦合并行计算机。这些系统非常适合超可靠的应用程序，因为它们内置的硬件冗余可以用来实现容错。该项目包括基础研究和应用研究两部分。基础研究部分主要研究多处理机系统故障诊断和容错路由问题的新模型和算法。应用研究部分涉及到容错多计算机系统实验试验台的开发。基于transpter的MIMD多计算机系统提供了测试平台硬件和底层操作系统。开发了专用软件，为试验台提供系统级容错框架、故障模拟器和数据收集工具。该试验台允许对在基础研究组件中开发的系统级容错机制进行实验评估，也可供从事多处理器系统容错工作的其他研究小组使用。

Recent developments have dramatically increased the importance of ultra-reliable parallel computing. First, parallel processing has moved from a specialized scientific computation tool to an instrument for general-purpose computing. At the same time, computers have begun to control virtually every aspect of human life and are, therefore, increasingly relied upon to provide continuous service. These trends make the development of general but practical techniques for fault-tolerant multiprocessing a necessity. This project focuses on loosely-coupled parallel computers. These systems are ideally suited for ultra-reliable applications due to their built-in hardware redundancy that can be used to achieve fault tolerance. The project has both basic and applied research components. The basic research component studies new models and algorithms for the problems of multiprocessor system fault diagnosis and fault-tolerant routing. The applied research component involves the development of an experimental testbed for fault-tolerant multicomputer systems. A Transputer-based MIMD multicomputer system provides the testbed hardware and low-level operating system. Special-purpose software is developed to provide a system-level fault tolerance framework, a fault simulator, and a data collection tool for the testbed. The testbed allows experimental evaluation of the system-level fault tolerance mechanisms developed in the basic research component and is also made available to other research groups working on multiprocessor system fault tolerance.