Monitors for Upset Detection in Computer Systems

Abstract : An upset is a deviation of system performance based on apriori knowledge of the system application task. There has heretofore been little experimental work reported in the literature regarding characterizations of upsets in computer systems. This problem was addressed in the research reported in this paper with two series of fault injection experiments on a representative microprocessor-based system. In the first series, the system executes a small machine language program; in the second series, a much larger software environment consisting of a Pascal interpreter running a general applications program is used. For each case, a broad spectrum of fault conditions is utilized for the fault injections. An extensive instrumentation complex has been developed to record microevent data associated with the upsets resulting from the injections. This data is used to characterize upsets in computer-based systems. It also serves as a benchmark for the performance of candidate upset monitors in the sense that coverage and latency figures for such devices can be deduced from this data. (Author)