Contributions on agreement in dynamic distributed systems

Résumé

This Ph.D. thesis studies the agreement problem in dynamic distributed systems by integrating both the classical fault-tolerance perspective and the more recent formalism based on evolving graphs. First, we developed a common framework that allows to analyze and compare models of dynamic distributed systems for eventual leader election. The framework extends a previous proposal by Baldoni et al. by including new dimensions and levels of dynamicity. Also, we extend the Time-Varying Graph (TVG) formalism by introducing the necessary timeliness assumptions and the minimal conditions to solve agreement problems. We provide a hierarchy of time-bounded, TVG-based, connectivity classes with increasingly stronger assumptions and specify an implementation of Terminating Reliable Broadcast for each class. Then we define an Omega failure detector, W, for the eventual leader election in dynamic distributed systems, together with a system model, , which is compatible with the timebounded TVG classes. We implement an algorithm that satisfy the properties of W in M. According to our common framework, M results to be weaker than the previous proposed dynamic distributed system models for eventual leader election. Additionally we use simulations to illustrate this fact and show that our leader election algorithm tolerates more general (i.e., dynamic) behaviors, and hence it is of application in a wider range of practical scenarios at the cost of a moderate overhead on stabilization times.