Bulk Insertions into R-Trees

A lot of recent work has focussed on bulk loading of data into multidimensional index structures in order to eeciently construct such structures for large datasets. Previous work on bulk loading data focussed at building index structures from scratch, while the problem of bulk insertions into existing index structures has been largely overlooked. In this paper, we address this new problem with particular focus on R-trees { which are an important class of index structures used widely in commercial database systems. We propose a new technique, which as opposed to the current technique of inserting data one by one, bulk-inserts entire new datasets into an active R-tree. This technique, called STLT (for Small-Tree-Large-Tree), considers the new dataset as an R-tree itself (small tree), identiies and prepares a suitable location in the original R-tree (large tree) for insertion, and lastly performs the insert of the small tree into the large tree. In this paper, we present an analytical model of STLT. Extensive experimental studies both on synthetic and real data sets from the Sequoia 2000 storage benchmark are also reported. These experiments not only compare STLT against the conventional technique, but also evaluate the suitability and limitations of STLT under diierent conditions, such as varying buuer sizes, ratio between existing and new data sizes, and skewedness of new data with respect to the whole spatial region. We nd that STLT does especially well (up to 80% better than the existing technique) for skewed datasets as well for large ratios of large tree to small tree data insertion sizes, while consistently outperforming the alternate technique in all other circumstances. Our experimental results also indicate that the quality of the resulting tree constructed by STLT in terms of query performance is acceptable and in most cases preferable over that created by the traditional tree insertion approach.