DELFIM: error detection by thin memory protection

Abstract

Memory protection has been widely recognized as an effective concurrent error detection method. Some of the studies made show that its effectiveness changes as the percentage of unused memory grows, since the probability of a wrong address falling on the unused zone also grows. In this paper, we follow that hint, and try to determine how better does the detection capability of protected memory get, when the protection boundaries are dynamically changed to encompass just the code of the function being executed. We implemented this method on a VME standard board, monitoring a commercial MC68000-based system. The system was tested by means of physical injection of about 25000 faults. Experimental results revealed that this mechanism can typically detect more than 80% of the faults, improving in about 5% the results from the traditional mechanisms. We also verified that only around 30% of the undetected faults induced a system failure.