Three-tier network database systems.

摘要

High-performance networked databases for contemporary data-intensive applications are created with powerful workstations/PCs interconnected via fast LAN and long-haul networks. Client-Server Databases CSDs) furnish promising response times for client-initiated transactions. However, even when client data-caching is used, the servers ultimately become the bottleneck as the number of concurrent client requests increases. In this thesis, we propose three-tier CSD architectures (3T-CSDs) that exploit user data access locality to realize a more scalable system configuration. Groups of clients that demonstrate similarities in their data access behavior are logically clustered together. Each such group of clients is handled by an Intermediate Cluster Manager (ICM) that acts as a duster-wide directory service and cache manager. Clients within the same cluster are now capable of sharing data among themselves without interacting the server(s). Through prototyping and experimentation, we show that the logical clustering of clients and the introduction of the ICM layer, significantly improves system scalability, and also improves transaction response times.; The logical client-clustering is optimal if the clients in each individual cluster have the maximum common data access probability. We show that the optimal client-clustering is NP-complete. To provide sub-optimal solution, we propose and evaluate two techniques used to create client clusters: a greedy algorithm and a genetic algorithm. We study the effect of the above two techniques in the performance of the 3T-CSDs. In addition, we examine the role of several preprocessing schemes used to reduce the volume of the input data to the clustering techniques.; When clients change their accessing behavior, the optimal quality of current client-clustering cannot be further guaranteed. In order to address this problem, we propose an on-line reclustering framework that realigns the clients according to their changing reference localities. The major contribution of our proposal lies in a change-detection algorithm that uses meta-data extracted from clients' data accesses. We evaluate the impact of employing a multi-featured change detection algorithm scheme in the 3T-CSD and experimentally investigate the various trade-offs involved.