Probabilistic Fault-tolerance of Distributed Services: A Paradigm for Dependable Applications

The paper deals with a study of probabilistic methods to manage the dependability of a networked distributed system, in the presence of inaccuracies and partial knowledge of system models pertaining to failures. A distributed networked system (DIS) that collects raw data from sensors deployed in the field and delivers a composite data to an end-user is faced with two types of uncertainties: at 'information level' due to the multi-modal nature of raw data collected from the environment, and at 'control level' due to the incompleteness in knowledge about the application model. These have a compounded effect on the quality of fault-tolerance exhibited by a DIS. Based on service-layer abstractions, the paper identifies application-oriented metrics to quantify the quality of information flowing through a DIS. Even with imperfect information, the paper demonstrates how the high-level quality metrics and control algorithms enable achieving a reasonable degree of fault-tolerance in a probabilistic manner. A case study of replicated web services is also described.