Experimental evaluation of the fail-silent behaviour in programs with consistency checks

Abstract

An important research topic deals with the investigation of whether a non-duplicated computer can be made fail-silent, since that behaviour is a-priori assumed in many algorithms. However, previous research has shown that in systems using a simple behaviour based error detection mechanism invisible to the programmer (e.g. memory protection), the percentage of fail-silent violations could be higher than 10%. Since the study of these errors has shown that they were mostly caused by pure data errors, we evaluate the effectiveness of software techniques capable of checking the semantics of the data, such as assertions, to detect these remaining errors. The results of injecting physical pin-level faults show that these tests can prevent about 40% of the fail-silent model violations that escape the simple hardware-based error detection techniques. In order to decouple the intrinsic limitations of the tests used from other factors that might affect its error detection capabilities, we evaluated a special class of software checks known for its high theoretical coverage: algorithm based fault tolerance (ABFT). The analysis of the remaining errors showed that most of them remained undetected due to short range control flow errors. When very simple software-based control flow checking was associated to the semantic tests, the target system, without any dedicated error detection hardware, behaved according to the fail-silent model for about 98% of all the faults injected.