A lightweight interface to predict communication delays using time series

Abstract

Chandra and Toueg first investigated the problem of solving consensus in asynchronous systems with unreliable failure detectors. In practice, such failure detectors work by suspecting consensus participants that do not respond in sufficient time, i.e., within a timeout (a maximum waiting time). However, calculating this timeout is difficult. In this way, many systems explore prediction techniques to dynamically forecast the communication delay used in timeout computation. In this paper, we are concerned with the timeout associated with failure detection in asynchronous systems in which processes may crash and links may lose messages. We propose a lightweight interface to transform a non-periodic sampling of the round-trip communication delay (rtt) observed by a pull-style failure detector in periodic sampling. By using this interface, we explore the use of time series predictors to dynamically forecast the rtt. As a result, we show that these predictors are more accurate than others commonly used on failure detectors. © Springer-Verlag Berlin Heidelberg 2003.