Towards Assessing Representativeness of Fault Injection-Generated Failure Data for Online Failure Prediction

Online Failure Prediction allows improving system dependability by foreseeing incoming failures at runtime, enabling mitigation actions to be taken in advance, though prediction systems´ learning and assessing is hard due to the scarcity of failure data. Realistic software fault injection has been identified as a valid solution for addressing the scarcity of failure data, as injecting software faults (the most occurring on computer systems) increases the probability of a system to fail, hence allowing the collection of failure-related data in short time. Moreover, realistic injection permits the emulation of software faults likely to exist in the target system after its deployment. However, besides the representativeness of the software faults injected is recognized as a necessary condition for generating valid failure data, studies on the representativeness of generated failure-related data has still not been addressed. In this work we present a preliminary study towards the assessment the representativeness of failure-related data by using G-SWFIT realistic software fault injection technique. We here address the definition of concepts and metrics for the representativeness estimation and assessment.