| The backend database system is often the performance bottleneck in multi-tier architectures. This is particularly true if there is a cluster of application servers while there is only a single database backend.;Existing approaches, however, have two limitations. First, they do not exploit the full capacity of the caches. Each application server has its own cache and frequently used objects will likely be cached in all caches, limiting the number of different objects and queries that can be cached. Furthermore, a query can only be served from the cache if previously the exact same query was posed.;In this thesis, we introduce Contquer, a distributed cooperative caching algorithm that uses a distributed caching architecture where each object is only cached at one application server and each application server has access to local and remote caches. Thus, the full capacity of all caches can be exploited. Furthermore, we optimize the query cache by exploiting the cache even if only part of a query can be served from the cache. For that we analyze the containment of queries within other queries. Contquer determines when a query can be fully or partially served from the cache, and automatically generates remainder queries to the database if necessary.;This thesis reports on the design and implementation of Contquer. It also conducts experiments that show that performance is improved considerably with the proposed algorithm. We conclude that the use of a distributed caching infrastructure and the ability to retrieve partial results from the cache improves performance in terms of hit-rate, throughput and latency.;A common approach to scale the database component is query result caching. The idea is to cache the results of a query submitted to the database in a cache. If the query is consequently requested again, the result can be retrieved from the cache instead of the query again being submitted to the database. Query caching can play a vital role in reducing latency by avoiding access to the database, and improving throughput by avoiding a database bottleneck. |
