查看文章

The abstraction of failure detectors came as an aid to the designers and programmers of dependable distributed applications. However, all implementations of failure detectors can make mistakes. In practice, timeout-based failure detector implementations improve their accuracy by adapting their timeout according to a stationary behavior assumption.

This paper analyzes the stationarity behavior of the communication delay, observed on a pull-style failure detector, by exploring the time series mathematical model. As a result, this paper shows that the behavior of the time series, instead of stationary, is non-stationary. From this result, we suggest the use of prediction models based on time series to improve the accuracy of failure detectors.