AReNA: Adaptive Distributed Catalog Infrastructure Based On Relevance Networks

Wide area applications (WAAs) utilize a WAN infrastructure (e.g., the Internet) to connect a federation of hundreds of servers with tens of thousands of clients. Earlier generations of WAA relied on Web accessible sources and the http protocol for data delivery. Recent developments such as the PlanetLab [8] testbed is now demonstrating an emerging class of dataand computeintensive wide area applications. One of the many challenging characteristics of WAA [14] involves the availability of several alternative data sources for a given query. This may be due to mirror sites, or more commonly, the overlap of content in specific application domains. For example, in scientific environments users may choose to access raw datasets, or they may be able to re-use the cached results of prior queries. Choosing among alternative data sources depends on access costs, coverage, rate of data delivery, outpt quality, etc. For example, a Performance Target (PT) [13] query optimizer may optimistically allow the selection of potentially noisy sources whose latencies range from excellent to poor since there is some expectation of meeting a performance target when there is no delay. A conservative strategy would avoid such a source since there was some expectation of missing the target. Access costs are know to exhibit transient behavior. The unpredictable behavior of a dynamic WAN [9, 11] results in a wide variability in access latency (end-to-end delay). Variation in access latency may depend on many factors including the topology of the client and the remote server, the network and server workloads, which are often affected by the time of day and day of week, and points of congestion between clients and servers [9]. While there are many