Section: New Results

Distributed algorithms for dynamic networks

Participants : Luciana Arantes [correspondent] , Olivier Marin, Sébastien Monnet, Franck Petit [correspondent] , Maria Potop-Butucaru, Pierre Sens, Julien Sopena, Raluca Diaconu, Ruijing Hu, Anissa Lamani, Sergey Legtchenko, Jonathan Lejeune, Karine Pires, Guthemberg Silvestre, Véronique Simon.

This objective aims to design distributed algorithms adapted to new large scale or dynamic distributed systems, such as mobile networks, sensor networks, P2P systems, Grids, Cloud environments, and robot networks. Efficiency in such demanding environments requires specialised protocols, providing features such as fault or heterogeneity tolerance, scalability, quality of service, and self-stabilization. Our approach covers the whole spectrum from theory to experimentation. We design algorithms, prove them correct, implement them, and evaluate them in simulation, using OMNeT++ or PeerSim, and on large-scale real platforms such as Grid'5000. The theory ensures that our solutions are correct and whenever possible optimal; experimental evidence is necessary to show that they are relevant and practical.

Within this thread, we have considered a number of specific applications, including massively multi-player on-line games (MMOGs) and peer certification.

Since 2008, we have obtained results both on fundamental aspects of distributed algorithms and on specific emerging large-scale applications.

We study various key topics of distributed algorithms: mutual exclusion, failure detection, data dissemination and data finding in large scale systems, self-stabilization and self-* services.

Mutual Exclusion and Failure Detection.

Mutual Exclusion and Fault Tolerance are two major basic building blocks in the design of distributed systems. Most of the current mutual exclusion algorithms are not suitable for modern distributed architectures because they are not scalable, they ignore the network topology, and they do not consider application quality of service constraints. Under the ANR Project MyCloud and the FSE Nu@age, we study locking algorithms fulfilling some QoS constraints often found in Cloud Computing [38] .

A classical way for a distributed system to tolerate failures is to detect them and then recover. It is now well recognized that the dominant factor in system unavailability lies in the failure detection phase. Regal has worked for many years on practical and theoretical aspects of failure detections and pioneered hierarchical scalable failure detectors. (Recent work by Leners et al published in SOSP 2011 uses our DSN 2003 paper as basis for performance comparison) Since 2008, we have studied the adaptation of failure detectors to dynamic networks. Following the model introduced in [18] , we have proposed new algorithms to detect crashes and Byzantine behaviors [32] .

These algorithms were designed as part of the ANR Project SHAMAN.

Self-Stabilization and Self-* Services.

We have also approached fault tolerance through self-stabilization. Self-stabilization is a versatile technique to design distributed algorithms that withstand transient faults. In particular, we have worked on the unison problem, (C. Boulinier, F. Petit, and V. Villain. Synchronous vs. asynchronous unison. Algorithmica, 51(1):61-80, 2008) i.e., the design of self-stabilizing algorithms to synchronize a distributed clock. As part of the ANR project SPADES, we have proposed several snap-stabilizing algorithms for the message forwarding problem that are optimal in terms of number of required buffers [36] . A snap-stabilizing algorithm is a self-stabilizing algorithm that stabilizes in 0 steps; in other words, such an algorithm always behaves according to its specification.

Finally, we have applied our expertise in distributed algorithms for dynamic and self-* systems in domains that at first glance seem quite far from the core expertise of the team, namely ad-hoc systems and swarms of mobile robots. In the latter, as part of ANR project R-Discover, we have studied various problems such as exploration [29] , and gathering [15] .

Dissemination and Data Finding in Large Scale Systems.

In the area of large-scale P2P networks, we have studied the problems of data dissemination and overlay maintenance, i.e., maintenance of a logical network built over the a P2P network. First, we have proposed efficient distributed algorithms to ensure data dissemination to a large set of nodes. Also, we have introduced a new method to compare dissemination algorithms over various topologies [35] .

MMOGs.

Peer-to-peer overlay networks can be used to build scalable infrastructures for MMOGs. Our work on MMOGs has primarily focused on the impact of latency constraints in dynamic distributed systems. In online P2P games, players are connected by a logical graph, implemented as an overlay network. Latency constraints imply that players that interact must remain close in the overlay, even when the mobility of players induces rapid changes in the graph.

We have also addressed problems related to cheating and arbitration. In a distributed system, certification of entities makes it possible to circumscribe malicious behavior, such as cheating in games. Certification requires the use of a trusted third party and is traditionally done centrally. At a large scale, however, centralized certification represents a bottleneck and a single point of attack or failure. We have proposed solutions based on distributed reputations to identify trusted nodes and use them as game referees to detect and prevent cheating [46] . Our method relies on previous work on the subject of trusted node collaboration to ensure reliable distributed certification (Erika Rosas, Olivier Marin and Xavier Bonnaire. CORPS: Building a Community Of Reputable PeerS in Distributed Hash Tables. The Computer Journal, 54(10):1721-1735(2011)).