Considering Workload Input Variations in Error Coverage Estimation

The effects of variations in the workload input when estimating error detection coverage using fault injection are investigated. Results from scanchain implemented fault injection experiments using the FIMBUL tool on the Thor microprocessor show that the estimated error non-coverage may vary by more than five percentage units for different workload input sequences. A methodology for predicting error coverage for a particular input sequence based on results from fault injection experiments with another input sequence is presented. The methodology is based on the fact that workload input variations alter the usage of sensitive data and cause different parts of the workload code to be executed different number of times. By using the results from fault injection experiments with a chosen input sequence, the error coverage factors for the different parts of the code and the data are calculated. The error coverage for a particular input sequence is then predicted by means of a weighted sum of these coverage factors. The weight factors are obtained by analysing the execution profile and data usage of the input sequence. Experimental results show that the methodology can identify input sequences with high, medium or low coverage although the accuracy of the predicted values is limited. The results show that the coverage of errors in the data cache is preferably predicted using data usage based prediction while the error coverage for the rest of the CPU is predicted more favourably using execution profile based prediction.