Optimal parallel I/O using replication

There has been a lot of interest in declustering spatial data for efficient parallel I/O. Declustering is used to distribute blocks of data among multiple devices, thus enabling parallel I/O access and reducing query response times. A strictly optimal declustering, or disk allocation, technique is the one that achieves optimal performance for all possible queries. We propose to use replication to reach optimal parallel I/O in multidisk/processor architectures. Replication is a well-known and effective solution for several problems in a database context, especially for availability and load balancing problems. Intelligent replication is considered. We especially focus on allocations based on latin squares and derive several nice properties for replication and declustering purposes. Three different replication strategies are proposed and evaluated. Using the proposed schemes, strict optimality is reached, even with a single replication, for all possible range queries and for several number of disks. We first show this for m-by-m grid on m disks, and then generalize it to any arbitrary a-by-b grids. We also show how to efficiently find optimal disk accesses for a given arbitrary query by storing minimal information.