Flexible Failure Detection with κ-FD

Many people rightly consider that failure detection should be provided as a generic distributed system service to be used to support fault-tolerance within many applications, in spite of possibly very diverse requirements. Reality is however different, as classical techniques are normally unable to cope with many distributed applications running simultaneously. To overcome this problem, we advocate a different approach to failure detection, whereby the failure detection service associates a real number to each process being monitored, representing the level of confidence that this process has crashed. In this paper, we present an adaptive failure detection algorithm based on the above principle. When compared experimentally with other recent failure detector implementations, we find that our algorithm performs significantly better with conservative failure detection (high accuracy, slow detection), and equally well with aggressive failure detection (quick detection, low accuracy). This comes in addition to a greater potential for flexibility offered by the model.