Using adaptive timeouts to achieve at-most-once message delivery

Abstract

We extend the algorithm of Liskov, Shrira, and Wroclawski for at-most-once message delivery so that it adapts dynamically to changes in message transmission time and degree of clock synchronization. The performance of their algorithm depends on its being supplied with a good estimate of the maximum message lifetime—the sum of the message delivery time and the difference in processor clock values between sender and recipient. We present two algorithms which are suitable for use in a system where the message lifetime is unknown or may change. Our extensions allow the automatic and continuous determination of a suitable value for the maximum lifetime. We prove that whenever the actual message lifetime is bounded, then our adaptive procedures converge to an accurate estimate of its true value. Our two algorithms make different assumptions about the behavior of the system and thus achieve different performance levels. Our formal statement of convergence is expressed in terms of the number of messages received, rather than time elapsed. We show that this formulation is necessary. Specifically, we prove that no method for estimating the lifetime can achieve convergence in a bounded amount of time.

This work was done while the first author was at the University of North Carolina at Chapel Hill. The work of the first and third authors was supported in part by NSF grant CCR-9010730 and an IBM Faculty Development Award.