On the Efficient Design and Testing of Dependable Systems Software

Modern computing systems that enable increasingly smart and complex applications permeate our daily lives. We strive for a fully connected and automated world to simplify our lives and increase comfort by offloading tasks to smart devices and systems. We have become dependent on the complex and ever growing ecosystem of software that drives the innovations of our smart technologies. With this dependence on complex software systems arises the question whether these systems are dependable, i.e., whether we can actually trust them to perform their intended functions. As software is developed by human beings, it must be expected to contain faults, and we need strategies and techniques to minimize both their number and the severity of their impact that scale with the increase in software complexity. Common approaches to achieve dependable operation include fault acceptance and fault avoidance strategies. The former gracefully handle faults when they occur during operation, e.g., by isolating and restarting faulty components, whereas the latter try to remove faults before system deployment, e.g., by applying correctness testing and software fault injection (SFI) techniques. On this background, this thesis aims at improving the efficiency of fault isolation for operating system kernel components, which are especially critical for dependable operation, as well as at improving the efficiency of dynamic testing activities to cope with the increasing complexity of software. Using the widely used Linux kernel, we demonstrate that partial fault isolation techniques for kernel software components can be enhanced with dynamic runtime profiles to strike a balance between the expected overheads imposed by the isolation mechanism and the achieved degree of isolation according to user requirements. With the increase in software complexity, comprehensive correctness and robustness assessments using testing and SFI require a substantially increasing number of individual tests whose execution requires a considerable amount of time. We study, considering different levels of the software stack, if modern parallel hardware can be employed to mitigate this increase. In particular, we demonstrate that SFI tests can benefit from parallel execution if such tests are carefully designed and conducted. We furthermore introduce a novel SFI framework to efficiently conduct such experiments. Moreover, we investigate if existing test suites for correctness testing can already benefit from parallel execution and provide an approach that offers a migration path for test suites that have not originally been designed for parallel execution