Overall Objectives
Scientific Foundations
Application Domains
New Results
Contracts and Grants with Industry
Other Grants and Activities

Section: New Results

Data Access with Autonomous Participants

Taking into account the autonomy of participants holds an important part in the evolution of data management, systems and applications, especially considering open systems such as internet. Autonomy is some kind of freedom left to participants which can be managed differently from one participant to another. It can take very different forms [16] . For example, it may mean “enter or leave the system at will as in P2P systems. Intuitively, the more participants are autonomous, the easier it is for a participant to integrate the system, but, the harder it is to manage the system. In this context, we have focused on two problems: satisfaction-based query allocation and peer representation for query routing.

Satisfaction-based Query Allocation

Participants : Philippe Lamarre, Jorge Quiane Ruiz, Patrick Valduriez.

This work is related to supporting participants' objectives in joining a system. Intuitively, in the field of open systems, it is the hope of achieving some objectives which motivates a peer to participate in a system. Obviously, different participants may have different objectives. Consequently, to integrate as many participants as possible, a system should not assume that all of them are interested in the same normative objective. Instead, a system should enable participants to act accordingly to their own private objectives. This approach has been studied in the field of query allocation.

In the context of dynamic distributed systems, with large numbers of heterogeneous, autonomous consumers and providers (the participants) query allocation is challenging because participants' interests may be contradictory. For example, a consumer would desire to receive results from a given provider but this provider would not desire to perform the query of such a consumer. In [24] , we defined a model to characterize the participants' satisfaction in the long run. We proposed a query allocation framework that takes into account participants' interests to satisfy them in the long-run. A particularity of our solution is that it dynamically makes the balance between providers' interests and consumers' interests taking into account their respective rewards. Experimental results show that our model enables a better evaluation of query allocation methods in these environments, and that our query allocation approach significantly outperforms baseline methods from both a satisfaction and performance point of view. This approach has been implemented and demonstrated [40] .

We have also studied satisfaction in the context of query replication. Indeed, multiple allocation of the same query can be used to solve problems related to distributed systems integrating autonomous participants: First, to restrict the effects of a provider failure, as a preventive measure, the system can allocate the query to many of them; Second, to deal with Byzantine participants, a query initiator is sometimes interested in allocating its query to many participants in order to compare their results. In both cases, the major problem is to find a good compromise between the number of replicas and a possible degradation of the system (performances and participants departure). We propose to take advantage of the huge capacity of adaptation of a satisfaction based approach to adapt the number of replicas, according to the allocation policy and participantsÕ global satisfaction. This method takes care of participants intentions and thus automatically adapts query allocation to their individual situations. Obtained results are under submission.

Peer Representation for Query Routing

Participants : Philippe Lamarre, Anthony Ventresque, Patrick Valduriez.

We consider unstructured P2P systems, the most general form. When querying, it is important to be able to characterize an answer with respect to the set of all possible answers, especially when they are obtained using a scoring function and not an exact match, as for example, in document search. Current Top-k algorithms find the k best answers to a given query, but they are generally quite costly since they require to visit all the peers that can be reached. However, intuitively, it should be possible to avoid some peers which would not improve the results. We followed this line assuming a peer is able to represent its neighbors in such a way that it can determine if a neighbor is likely or not to improve the already obtained results. In [43] we studied the properties such a representation should satisfy 1 - to warranty the result of a Top-k query while avoiding some participants, and 2 - to be easily maintained. We then proposed a representation using semantic vectors and we proved it satisfies these properties. We presently work on algorithms to maintain such representation within a dynamic system and on related search algorithms.


Logo Inria