R-tree-based data migration and self-tuning strategies in shared-nothing spatial databases

Abstract

In order to provide fast and timely answers to queries in the context of spatial databases and GIS, we present our solution for effective data migration and tuning strategies in shared-nothing parallel spatial databases. Our purpose is to improve the performance of the indexes. Our approach has the following features. First, our scheme is self-tuning, dynamic as well as query-centric and it can adapt to dynamically changing user access patterns. Second, a global distributed R-tree-based indexing method is employed to facilitate effective data migration. Third, unlike traditional partitioning strategies where each processing element (PE) contains data from a single region of space, we allow each PE to store data from multiple and disjoint regions. This minimizes overlap in regions as well as coverage. We implemented the proposed scheme and conducted an extensive performance study on Fujitsu's AP3000 machine with 32 workstations using real datasets. Our experimental results show that our load-balancing strategy can distribute the load effectively across the PEs in the system, thereby reducing response times of incoming queries.