Robust Fault Tolerance for Computations in Linear Algebra and Signal Processing

线性代数和信号处理计算的鲁棒容错能力

• 批准号: 9628786
• 负责人: Daniel Boley
• 金额: $ 5.99万
• 依托单位: University of Minnesota-Twin Cities
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 1996
• 资助国家: 美国
• 起止时间: 1996-10-01 至 1999-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=9628786&HistoricalAwards=false
• 关键词: Robust; Fault; Tolerance; Computations; Linear

The objective of this project is to develop and test fault tolerant algorithms for numerical computation in various domains, including sparse matrices, recursive filtering and signal processing. The methods used are a combination of checks made during the computation based on preconditioned checksums and thresholds, and a-posteriori checks made on the final computed results based on backward error assertions (BEAs). In spite of the presence of rounding errors, BEAs allow very tight, cheap, validation of the final results independent of the specifics of the implementation or sources of error, and checksums and thresholds supplement the BEAs by providing checks during the computation. These methods will be extended beyond prototype examples to more realistic large computations, demonstrating their scalability and applicability to large domains.***

这个项目的目标是开发和测试用于不同领域的数值计算的容错算法，包括稀疏矩阵、递归滤波和信号处理。所使用的方法是基于预条件校验和和阈值在计算期间进行的检查和基于后向错误断言(BEA)对最终计算结果进行的后验检查的组合。尽管存在舍入误差，但BEAS允许对最终结果进行非常紧密、廉价的验证，而不受实施细节或误差来源的影响，而校验和和阈值通过在计算期间提供检查来补充BEAS。这些方法将从原型示例扩展到更真实的大型计算，展示它们对大型领域的可扩展性和适用性。*