Global Memory Management in Client-Server DBMS Architectures

Earlier performance studies of client-server database systems have investigated algorithms for caching locks and data at client worhxtations to reafuce latency and offload the server. These studies have been restricted to algorithms in which d&abase pages that were not in the local client buffer pool or the server buffer pool were read in from disk. In this paper we investigate a technique that allows client page requests to be serviced by other clients, thus treating the entire system as a single memory hierarchy. We also present techniques for efficiently exploiting this global memory hierarchy by reducing the replication of pages between client and server buffer pools. Global memory management algorithms that employ various combinalions of these techniques are then described, and the performance tradeoffs among the algorithms we investigated under a range of workloads and system conjiguratio~ using a simulation model.