An approach to generalising the self-repair of overlay networks

Overlay networks are emerging as the technology of choice for deploying new large-scale distributed services and applications. Despite being such a key technology, overlays have no coherent strategy for failure recovery, employing a plethora of ad-hoc solutions. This thesis surveys the state-of-the-art in overlay network and fault-tolerance research, examining the shortcomings of present approaches. It then proposes SONAR, a highly configurable and adaptive distributed service, to provide generalised fault-tolerance to a wide range of overlay networks. The design of SONAR is discussed, and its unique API is presented in detail—an API enabling genericity of SONAR while permitting expressiveness of an overlay’s specific traits. The heart of the proposal is then presented and discussed; a new kind of distributed repair protocol, which leverages distributed consensus to provide a strong base from which multiple kinds of repair can be dynamically selected and safely enacted at runtime. This in turn enables adaptive repair, which can intelligently decide on the best course of action for each failure. The proposed service is then evaluated in its ability to support several major and varied overlays, showing that the approach taken is viable despite its genericity, and demonstrating the powerful capabilities of an adaptive repair mechanism. The outcome of this proposal is a re-usable service which can be employed by overlay developers to support repair in their overlay, successfully achieving a clean separation of concerns, and leaving the developer to focus on the functional aspects of their design.