On the Feasibility of Unstructured Peer-to-Peer Information Retrieval

We consider the feasibility of web-scale search in an unstructured peer-to-peer network. Since the network is unstructured, any such search is probabilistic in nature. We therefore adopt a probably approximately correct (PAC) search framework. The accuracy of such a search is defined by the overlap between the set of documents retrieved by a PAC search and the set of documents retrieved by an exhaustive (deterministic) search of the network. For an accuracy of 90%, we theoretically determine the number of nodes each query must be sent to for three distributions of documents in the network, namely uniform, proportional and square root. We assume that the query distribution follows a power law and investigate how performance is affected by the scale factor. For various configurations, we estimate the global and local network traffic induced by the search. For a network of II. million nodes, a query rate of 1000 queries per second, and assuming each node is capable of indexing 0.1% of the collection, our analysis indicates that the network traffic is less that 0.07% of global internet traffic.