Two-ways Adaptive Failure Detection with the φ-Failure Detector

It is widely recognized that distributed systems would greatly benefit from the availability of a generic failure detection service. Such a service can however prove useful only if it can adapt simultaneously to changing network conditions and conflicting application requirements. This paper presents a novel approach to adaptive failure detectors, called φ-failure detectors, which dynamically adapts to application requirements, as well as network conditions. The key idea is as follows. Traditionally, failure detectors maintain a set of suspected processes. The information is hence of boolean nature, that is, some process p is suspected if and only if it belongs to this set. In contrast, a φ-failure detector associates a value φp to every known process p. The value φp increases according to a normalized scale which represents the degree of confidence that process p has crashed. The scale is dynamically adapted from the current network conditions, and each application can trigger suspicions according to a threshold which corresponds to its own requirements. We describe a possible implementation for such a service, although some specific questions remain open where this work is still in progress.