LoT: Dynamic Declustering of TSB-Tree Nodes for Parallel Access to Temporal Data

In this paper, we consider the problem of exploiting I/O parallelism for efficient access to transaction-time temporal databases. As temporal databases maintain historical versions of records in addition to current ones, we consider range queries in both time dimension and key dimension. Multiple disks can be used to read sets of disk blocks in parallel, thereby improving the performance of such queries substantially. The problem is to find an optimal declustering algorithm for spreading record versions across disks. The solution depends on the index structure used. We have adopted the time split B-tree, as it provides efficient support for time range and key range queries. Our declustering method coined LoT (Local Balancing for TSB-trees) aims to decluster runs of logically consecutive leaf nodes of a TSBtree onto separate disks. The method is dynamic in the sense that it computes the disk address of a new node at its creation time, based on the disk addresses of the nodes in its neighborhood. LoT is an extension of the local balancing algorithm presented in [SL91]. It considers different sets of disks for historical and current nodes, and uses a two-dimensional distance metric between TSB-tree leaf nodes. As historical nodes of TSB-trees are no longer subject to splits, the coordinates of new nodes in the time-key space are restricted. This is exploited in LoT for achieving good declustering for both time range and key range queries. We derive performance guarantees for LoT in terms of the speedup for range queries. Simulation results show the response time speedup of LoT compared to a scheme that assigns nodes to disks in a random manner.