On the Use of Assembly Code Metrics for Error Coverage Prediction

Abstract

In this thesis we present two prediction techniques for estimating the error coverage of target programs stimulated with different inputs. Preliminarily, we investigate the effects of the inputs on the failure distribution of the target programs using fault injection experiments. From this study, we could find a linear correlation between the length of the input and the error coverage. This result allows us to develop a linear regression model which is one of the prediction techniques that we adopt. As this correlation may not exist in other target programs, in the second technique called instruction-based prediction we propose an approach to predict the error coverage for an input using fault injection results of other inputs known as base points. In order to choose the base points, instruction-based prediction technique profiles the program through a set of metrics defined at the assembly code. Those metrics are used to feed a statistical technique that helps us select the more suitable inputs for the prediction. We also investigate the failure distributions of programs enhanced with the triple time redundancy execution with forward recovery (TTR-FR). From the results of the failure distributions, we observe that the non-covered failure is reduced to on the average around 1.2% for all TTR-FR execution flows which has a minor correlation to input length as analyzed by linear regression equation.