QoS of timeout-based self-tuned failure detectors: the effects of the communication delay predictor and the safety margin

Unreliable failure detectors have been an important abstraction to build dependable distributed applications over asynchronous distributed systems subject to faults. Their implementations are commonly based on timeouts to ensure algorithm termination. However, for systems built on the Internet, it is hard to estimate this time value due to traffic variations. Thus, different types of predictors have been used to model this behavior and make predictions of delays. In order to increase the quality of service (QoS), self-tuned failure detectors dynamically adapt their timeouts to the communication delay behavior added of a safety margin. In this paper, we evaluate the QoS of a failure detector for different combinations of communication delay predictors and safety margins. As the results show, to improve the QoS, one must consider the relation between the pair predictor/margin, instead of each one separately. Furthermore, performance and accuracy requirements should be considered for a suitable relationship.