Experiences with Load Balancing and Caching for Semantic Web Applications

In our work we consider applications which generate queries w.r.t. many different knowledge bases. We presuppose that for a particular KB there exists many possible query servers. In order to successfully build applications that exploit these KB servers, an appropriate middleware is required. In particular, if there are many servers for a specific KB, the middleware is responsible for managing request dispatching and load balancing. Load balancing must be accompanied by middleware-side caching in order to reduce network latency. In our view the KB servers we consider are managed by different organizations. Therefore, DL applications used in some company need some gateway inference server that provides local caching (in the intranet) to: (i) reduce external communication and (ii) avoid repetitive external server access operations in case multiple intranet applications pose the same queries. In our case study we investigate a server for answering OWL-QL− queries1. This server (called RacerManager) acts as a proxy that delegates queries to back-end DL reasoners (RacerPro servers) that manage the KB mentioned in the query and load KBs on demand. Compared to previous versions, the functionality of RacerManager has been substantially enhanced. We address the problems of load balancing and caching strategies in order to exploit previous query results (possibly produced by different users of the local site). Caching is investigated in the presence of incrementally answered OWL-QL− queries. In addition, the effects of concurrent query executions on multiple (external) inference servers and corresponding transmissions of multiple partial result sets for queries are studied.