Due to their fundamental role in the design of faulttolerant distributed systems, consensus protocols have been widely studied. Most of the research in this area has focused on providing ways for circumventing the impossibility of reaching consensus on a purely asynchronous system subject to failures. Of particular interest are the indulgent consensus protocols based upon weak failure detection oracles. Following the first works that were more concerned with the correctness of such protocols, performance issues related to them are now a topic that has gained considerable attention. In particular, a few studies have been conducted to analyze the impact that the quality of service of the underlying failure detection oracle has on the performance of consensus protocols. To achieve better performance, adaptive failure detectors have been proposed. Also, slowness oracles have been proposed to allow consensus protocols to adapt themselves to the changing conditions of the environment, enhancing their performance when there are substantial changes on the load to which the system is exposed. In this paper we further investigate the use of these oracles to design efficient consensus services. In particular, we provide efficient and robust implementations of slowness oracles based on techniques that have been previously used to implement adaptive failure detection oracles. Our experiments on a widearea distributed system show that by using a slowness oracle that is well matched with a failure detection oracle, one can achieve performance as much as 53.5% better than the alternative that does not use a slowness oracle.
consensus protocols; asynchronous distributed systems; adaptive protocols; slowness oracles; unreliable failure detectors; predictors
