Software Implemented Fault Tolerance in Multicomputers

这项提议的主要目标是开发一种提高并行计算机可靠性的综合方法。对于本地(节点)级硬件故障的检测和诊断，该方法将结合离线内建自测试、在线并发检查以及实现基于路径和算法的检查的软件技术。这项提议的目的是开发一种系统的方法，用于生成和检查应用程序的执行情况。为了自动化核查过程和在程序中放置支票，将开发一种名为Verify的互动工具。为了估计这些技术提供的覆盖率，将创建一个软件故障注入工具。将开发高效、快速的重构算法，以避免单个和多个故障节点和链路。将创建一个用户透明的软件层来实现所提出的算法。这将被写在现有的系统软件之上，这些软件驻留在并行计算系统的每个节点上。面向软件对象的故障注入工具(SOFIT)将被移植到Ncube和Tandem SeverNet上。恢复时间开销将在程序执行时SOFIT注入Ncube和SeverNet的节点和链路的单个和多个故障的情况下进行估计。将进行可靠性评估研究。这项研究的目的是证明这种综合方法可以成功地用于将Ncube和SeverNet的可靠性提高一个数量级。

The main objective of this proposal is to develop an integrated approach for increasing the dependability of parallel computers. For detection and diagnosis of hardware faults at the local (node) level, this approach will combine off-line built-in self testing, on-line concurrent checking, and software techniques implementing path and algorithmic-based checks. The aim of this proposal is to develop a systematic method for generating and placing checks into the implementation of an application. For automating the verification process and placement of checks in programs, an interactive tool, VERIFY, will be developed. For estimating the coverage factor provided by these techniques, a software fault-injection tool will be created. Efficient and fast reconfiguration algorithms will be developed for avoiding single and multiple faulty nodes and links. A user transparent software layer will be created for implementing the proposed algorithms. This will be written on top of the existing system software that residents in every node of parallel computing systems. The Software Object-Oriented Fault-Injection Tool (SOFIT) will be ported on the nCUBE and Tandem SeverNet. The recovery time overhead will be estimated in the presence of single and multiple faults injected by SOFIT into the nodes and links of the nCUBE and SeverNet when programs are being executed. A dependability evaluation study will be performed. The objective of this research is to show that this integrated approach can be used successfully for improving the reliability of the nCUBE and SeverNet by an order of magnitude.