On the Use of Event Logs for the Analysis of System Failures

The level of trust on log-based dependability characterization of complex distributed systems, is biased by the ability at identifying failure data from event logs collected across different locations, which is a challenging issue. Several factors compromise the ability at identifying failure data, such as the accuracy of the logging mechanism to detect occur- ring errors, the effectiveness of the infrastructure that is adopted to manage failure data, correlation phenomena among the entries in the log. The focus of the thesis is to evaluate the accuracy of current logging mechanisms at reporting failures, and to develop novel techniques to make event logs effective to infer failure data. Techniques involve production, collection, and correlation of the failure data in the log to support accurate system dependability characterization. Despite the existence of a large number of studies proposing dependability measurements and models starting from collected logs, this is the first contribution that investigates and improves the accuracy of data sources and procedures, which are commonly adopted to determine the failure-related information used to conduct a dependability study.