Decentralized Fault Management for Service Dependability in Ubiquitous Networks

Obtaining reliable operation of end-user services in future ubiquitous networking environments is challenging. Faults occur and heterogeneous networks make it difficult to deploy network wide fault management mechanisms. This PhD lecture presents a study on the options an end-node has to mitigate faults in the end-to-end path. Such end-node driven fault management aims to make use of the diversity in the networking environment (various technologies, providers and operational characteristics) to provide improved resiliency without explicit network support. The resulting challenges are: i) unobservable and incomplete network state information, ii) unreliable observations, and iii) dynamic environments calling for adaptation in the fault management process. In the study, focus is on potential gains in the interaction between the components of Observation, Diagnosis, Decision and Remediation. Initially, it is shown how diagnosis robustness may be improved towards unreliable observations by adopting multiple cross-layer observations and decoding measurement uncertainty information. Next, a case network scenario study is introduced for a time constrained end-user service. The scenario enables insights on the impact of unavoidable diagnosis imperfections on service reliability. Also, it is studied to what extent good remediation decisions may be applied to mitigate such imperfections. For this purpose a light-weight decision policy evaluation model is proposed and verified in a system level simulation model. Some of the main findings are: i) certain imperfection trade-off settings of the Diagnosis component lead to worse end-user service reliability than if no fault management is conducted, ii) using end-user service state information in the decision process can help improve service reliability by ignoring imperfect diagnosis in non-critical states, and iii) imperfections of complex diagnosis mechanisms can in a parsimonious manner be represented in the proposed policy evaluation model to identify the best diagnosis trade-off setting. Finally, an outlook is provided to assess how the policy evaluation model may be re-configured dynamically to changes in the networking environment.