A Dataset of Linux Failure Data for Dependability Evaluation and Improvement

Software systems are now used to execute critical tasks on a daily basis. As a result, unhandled or uncontrolled failures at runtime may lead to non-negligible risks or losses. To mitigate this, considerable effort and resources have been dedicated to assessing and improving the dependability of such systems. However, researching novel techniques to develop more dependable systems requires access to rich and detailed data. As data from real systems are not typically available, researchers often look for alternative processes, such as fault injection, to generate realistic synthetic data. As this requires considerable effort and expertise, researchers frequently rely on outdated datasets or develop simplified processes to collect data, eventually compromising the validation and development of their methods. This paper presents, discusses, and makes available a large failure dataset collected from an up-to-date Linux kernel through fault injection. It provides a detailed characterization of the target system by continuously monitoring hundreds of system metrics and various system logs throughout the experiments. Ultimately, the goal is to provide a reliable, well-defined, and properly generated dataset that can be used to research techniques to support the development of more dependable systems.