The risk of data corruption in microprocessor-based systems

Most microprocessor-based systems provide little or no checking for failures of the microprocessor itself. These systems rely on the high reliability of VLSI and the inherent run-time checks (e.g., virtual memory protection and illegal instruction exceptions) to ensure data integrity. The authors present a model for analyzing the vulnerability of systems to undetected data corruptions. The model estimates expected corruption rates based on commercial reliability studies, field experience with fault-tolerant systems, and academic fault-injection studies. They find that without lockstep, an undetected data corruption may occur nearly once per month in a population of 10,000 processors. Given the available data, however, it is difficult to provide tight bounds for the frequency of data integrity violations and the authors' current estimates of the data corruption rate vary by more than two orders of magnitude. Further study on transient failures and the impact of faults is required to narrow the range of the predictions and reduce the overall risk of undetected data corruptions.