Section: Scientific Foundations
Introduction
Economic activities and human lives are now heavily dependent on distributed systems and applications. When computing resources and stored data can be affected by the occurrence of failures, dependability becomes a crucial issue.
When a low level of dynamicity (also called churn) is assumed or when the system size is rather small, a process involved in a distributed computation may know and observe all the other participants. Distributed applications often rely on the identification of such sets of interacting entities. These small sets are either called groups, clusters, collections, neighborhoods, spheres, or communities according to the criteria that define the membership. The adopted criteria may for instance reflect the fact that its members are administrated by a unique person, that they share a unique security policy, that they are located in closed physical places, that they need to be strongly synchronized, that they cooperate together, or that they share mutual interests. When all the participants can share a common knowledge of the group of interacting processes, various fundamental problems (related to observation and synchronization) can be solved easily. Adaptive algorithms can be proposed to detect a modification of the whole execution context and react globally to this modification (reconfiguration, execution of another code, ...). In particular, to cope with the dynamic evolution of a distributed system, the Group paradigm (and the associated concept of membership service) allows to efficiently address dependability issues. Solutions to agreement problems (such as the consensus problem) can be used as basic building blocks for designing solutions to higher level protocols that are in charge of maintaining global properties at the group level despite the occurrence of faults within the group. Due to the increasing adversity of the system (asynchrony and failures), the design of efficient solutions that are simple to deploy and easy to adapt remains a difficult issue.
When the system has a very high level of churn, implementing a global observation mechanism that allows to reconfigure the whole system in a single step is no more realistic. Only local observations and progressive adaptations to changes can be performed on cohesive subsets of nodes. Such a radical gap on the scale and dynamicity of systems militates in favor of a paradigm shift for designing solutions to the problems raised by these new systems. Several partial and inconsistent views of the system may coexist (each participant may have its own view). All classical distributed computing problems (for example, dependability issues, communication problems, resource allocation, and data management) require new solutions that address these challenges in the new settings. In the context of large-scale distributed and dynamic systems, interaction with unknown entities becomes an unavoidable habit despite the induced risk. In this general context, we consider mainly reputation mechanisms in P2P systems and privacy protection issues.
