Distributed network querying with bounded approximate caching

As networks continue to grow in size and complexity, distributed network monitoring and resource querying are becoming increasingly difficult. Our aim is to design, build, and evaluate a scalable infrastructure for answering queries over distributed measurements, at reduced costs (in terms of both network traffic and query latency) while maintaining required precision. In this infrastructure, each network node owns a set of numerical measurements and actively maintains bounds on these values cached at other nodes. We can answer queries approximately, using bounds from nearby caches to avoid contacting the owners directly. We focus on developing efficient and scalable techniques to place, locate, and manage bounded approximate caches across a large network. We have developed two approaches: One uses a recursive partitioning of the network space to place caches in a static, controlled manner, while the other uses a locality-aware distributed hash table to place caches in a dynamic and decentralized manner. In this paper, we focus on the latter approach. Experiments over a large-scale emulated network show that our techniques are very effective in reducing query costs while generating an acceptable amount of background traffic; they are also able to exploit various forms of locality that are naturally present in queries, and adapt to volatility of measurements.