BB-Tree: A practical and efficient main-memory index structure for multidimensional workloads

We present the BB-Tree, a fast and space-efficient index structure for processing multidimensional read/write workloads in main memory. The BB-Tree uses a k-ary search tree for pruning and searching while keeping all data in leaf nodes. It linearizes the inner search tree and manages it in a cache-optimized array, creating the need for occasional re-organizations when data changes. To reduce the frequency of such re-organizations, the BB-Tree introduces a novel architecture for leaf nodes, called bubble buckets, which can automatically morph between different representations depending on their fill degree and are thus able to buffer a large number of insertions or deletions in-place. We compare the BB-Tree to scanning, main-memory variants of the R∗-tree, the kd-tree, and the VA-file, and the recent PHtree using different multidimensional workloads over real and synthetic data sets. The BB-Tree is the fastest access method for range queries up to a selectivity of around 20% (after which it is only beaten by scanning), the fastest method in read/write workloads, and achieves an exact-match query performance similar to that of the best point access method. In addition, it is the most space-efficient method of all considered index structures. We also describe a parallel range query operator and show that it scales with the number of physical cores.