Section: New Results

Massively Distributed Data Management Systems

Our work on the AMADA platform has shown how the different sub-systems of a popular cloud platform (namely, Amazon Web Services, or AWS in short) can be harnessed to build scalable stores and query evaluation engines for XML and RDF data. In [23] , we propose and compare several storage and indexing strategies within AWS, and show that they help reduce not only query evaluation time but also the monetary costs associated to the exploitation of the AWS-based store, since the index helps direct queries only to the subsets of the data likely to have results for the query. Thus, the total effort (and the costs charged by AWS) in relation to the processing of a given query are reduced. A similar study focused mostly on RDF data management appears as a book chapter [40] . More information can be found at http://cloak.saclay.inria.fr/research/amada/ .

Semantic Web data collections, that is, RDF graphs, may be very voluminous since RDF natively enables connections between different RDF databases (which may have been produced independently and in ignorance of each other) through the usage of common URIs (resource identifiers) in two or more databases. To scale up to such large volumes, we have developed CliqueSquare, a novel platform for storing and querying RDF graphs in a MapReduce-based architecture such as Hadoop. We have described the storage and query algorithm in [34] . Our analysis of existing frameworks and algorithms for managing large RDF graphs in a highly distributed environment has lead to the tutorial [27] .

Large-scale distributed processing of complex data was considered from a different perspective in our Delta project. Here, we considered the setting where one data source publishes new data items at a very high rate, and numerous clients subscribe to some of the updates by means of queries that must be matched by the published items. In this setting, the source may quickly become the bottleneck due to limitations in its capacity to match the published item against the subscription and/or to send the matching updates. We propose a fully automated approach for distributing the data dissemination effort across the network of subscribers, by identifying some which act as secondary data sources for others, in a peer-to-peer fashion. This distributed dissemination network is chosen so as to optimize a combination of overall dissemination costs and data propagation latency; since the space of options has daunting complexity, approximate algorithms involving Binary Integer Programming techniques were proposed in [20] , [37] , [42] , and concluded in the PhD thesis of A. Katsifodimos [11] .