Definition and properties of accrual failure detectors : an overview

Ensuring fast and accurate failure detection is a fundamental issue for building efficient fault-tolerant distributed systems. In an effort to make fault-tolerant applications easier to implement, we are trying to provide failure detection as a generic Internet service, similar to what was done very successfully with NTP (network time protocol) for clock synchronization. To do so, we must revisit the interaction model between the failure detection service and the applications it serves. We have defined a novel failure detector abstraction called accrual failure detectors. The main different with traditional failure detectors is that, instead of outputing information of a binary nature (trust or suspect), accrual failure detectors output a level of confidence. Last year, we presented preliminary results on the implementation of a failure detector based on this principle [7]. In this presentation, we will show our recent advances on this issue. In particular, we will present the definition of accrual failure detectors. Such a definition is necessary for clearly establishing the relation with the already abundant work on failure detectors.