Bookshelf

Considering the growing importance of distributed systems in today’s society, such as for cloud computing applications, the need for ensuring the systems’ dependability intensifies likewise. In the context of complex, fast-evolving distributed systems, the approach of software fault injection for their experimental dependability assessment does not seem to be unfolded to its full potential yet.

This work presents a structured method to derive a software fault injection campaign from a user-provided dependability model. The deterministically derived campaign aims to test for synergistic effects and therefore identifies all combinations of as many concurrently tolerable faults as possible, while also optimizing for efficiency. An implemented tool can programmatically derive the campaign and coordinate its exercise via user-provided executables, which, in turn, implement the specifics of the actual dependability assessment. For the case study on OpenStack, its fault tolerance mechanisms are elaborated on in detail and consolidated in a dependability model, represented as a fault tree. Practical evaluations of a fully virtualized high availability OpenStack show that its setup, as well as its checkpointing and rollback are challenging. The efforts finally yielded a setup which allowed to implement the completely automated exercise of the campaign. On top of OpenStack, the application Tahoe Least Authority File Store is used as a workload, and to measure the performance degradation while exercising the campaign. The high availability OpenStack setup does not reveal any outstandingly critical components, and Tahoe-LAFS experiences performance degradations between factor ~1 and ~4, with an average of ~1.8.