Iterative-improvement-based declustering heuristics for multi-disk databases
Information Systems
| Hi-index | 0.00 |
Several methods for declustering spatial (i.e, multi-dimensional) data on a set of M disks have been proposed in recent years, in order to reduce response time for large range queries and to increase level of concurrency for small range and point queries. Some of these methods provide dynamic management of data only within disks, by using M parallel independent (PI) organizations and a static mapping of regions of space to disks that guarantees correct addressability. On the other hand, so-called multiplexed (MX) parallel structures, such as MX R-trees, dynamically decluster a single-disk spatial structure (e.g., an R-tree) without using any data-independent mapping function. The contribution of this paper to the development of efficient parallel spatial structures is twofold. First, we introduce a new declustering method, called MC^2, based on a dynamic mapping function that assigns new regions of space to disks by using Multiple error Correcting Codes. Differently from previous methods based on static mapping functions, MC^2 requires no global reorganization, because of the way splits of data regions are dealt with, and can therefore deal also with (highly) dynamic data sets and non stationary data distributions. As a second contribution, we show how dynamic management of data across disks can be achieved even with a PI organization without losing global addressing capabilities. To this end, a new organization based on the grid file concept, namely PI-GF (`Parallel Independent Grid Files''), is introduced and its behavior formally specified.