Dynamically Reconfigurable Filtering Architectures

Abstract

Distributed R-trees (DR-trees) are appealing infrastructures for implementing range queries, content based filtering or k-NN structures since they inherit the features of R-trees such as logarithmic height, bounded number of neighbors and balanced shape. Interestingly, the mapping between the DR-tree logical nodes and the physical nodes has not yet received sufficient attention. In previous works, this mapping was naively defined either by the order physical nodes join/leave the system or by their semantics. Therefore, an important gap in terms of load and latency can be observed while comparing the theoretical work and the simulation/experimental results. This gap is partially due to the placement of virtual nodes. A naive placement that totally ignores the heterogeneity of the network may generate an unbalanced load of the physical system. In order to improve the overall system performances, this paper proposes mechanisms for placement and dynamic migration of virtual nodes that balances the load of the network without modifying the DR-tree virtual structure. That is, we reduce the gap between the theoretical results and the practical ones by injecting (at the middleware level) placement and migration strategies for virtual nodes that directly exploit the physical characteristics of the network. Extensive simulation results show that significant performance gain can be obtained with our mechanisms. Moreover, due to its generality, our approach can be easily extended to other overlays or P2P applications (e.g. multi-layer overlays or efficient P2P streaming).