Reversible Phase Transitions in a Structured Overlay Network with Churn

Abstract

Abstract. Distributed applications break down when the underlying system has too many node or communication failures. In this paper, we propose a general approach to building distributed applications that lets them survive hostile conditions such as these failures. We extend an existing Structured Overlay Network (SON) that hosts a transactional replicated key/value store to be Reversible, i.e., it is able to regain its original functionality as the environment hostility recedes. For this paper we consider the environment hostility to be measured by the Churn parameter, i.e., the rate of node turnover (nodes failing and being replaced by new correct nodes). In order to describe the qualitative behavior of the SON at high churn, we introduce the concept of Phase of the SON. All nodes in a phase exhibit the same qualitative properties, which are di↵erent for the nodes in di↵erent phases. We demonstrate the existence of Phase Transitions (i.e., a significant fraction of nodes changes phase) as churn varies and show that our concept of phase is analogous to the macroscopic phase of physical systems. We empirically identify the Critical Points (i.e., when there exists more than one phase simultaneously in significant fractions of the system) observed in our experiments. We propose an API to allow the application layer to be informed about the current phase of a node. We analyze how the application layer can use this knowledge for self-adaptation, self-optimization and achieve reversibility in the application-level semantics.