Applied research - search optimization for p2p networks

Search Optimization in the Distributed Networks.: by S.Osokin.

Abstract
Superpeers (also sometimes called ultrapeers) are generally considered to be a promising approach to content search improvement in the distributed networks. This study analyzes the performance of several superpeer architectures from the practical standpoint, comparing the different architectures and superpeer role assignment algorithms when the goal is to to maximize the search query reach in a peer-to-peer network. The study also investigates the performance of the 'Local Indices' mutual index caching architecture from the same standpoint, and suggests an optimal caching approach that tries to maximize the peer-to-peer search performance by distributing the superpeer functionality across all network nodes. This approach can significantly improve the query reach in comparison with the superpeer architectures, since it allows all network hosts to maximize their contribution to the overall network resource pool, potentially approaching the theoretical limits of the peer-to-peer network search performance. The study shows that the search performance of this hybrid approach is directly tied to the average host bandwidth and to its hardware (RAM and CPU) resources, as opposed to the normal superpeer case, when the search performance is tied to the bandwidth and resources of the 'best' network hosts. Thus the suggested algorithm does not require careful superpeer selection and does not present the high-profile superpeer targets to the attacker, which makes it less fragile than the conventional superpeer architectures.

...and Conclusion
Search performance is one of the most important characteristics of peer-to-peer networks. This study analyzes the performance of the peer-to-peer networks with a non-guaranteed search that represent the majority of the most popular peer-to-peer systems today. The study shows that for the superpeer-based architectures the best search performance can be achieved by using the massively redundant superpeer clusters, and that today's openly discussed peer-to-peer architectures just scratch the surface of the potential search performance of the peer-to-peer systems.

After that, the study analyzes the search performance of the mutual index caching architecture and shows that this architecture can achieve the best theoretically possible search performance for the non-guaranteed search networks, since it can fully utilize the available resources of all network hosts, whereas the superpeer architectures utilize only the resources available on the most powerful hosts. For the practically interesting case of the ad hoc network with a finite host session time the study suggests the optimal volume of the mutually cached data on the host. It also proposes the simple fully decentralized network bandwidth management algorithm for the hosts in this network, showing that with the network parameters typical for today's file-sharing systems this algorithm outperforms the massively redundant superpeer clusters, and that this performance gap will continue to widen as the average host session time will grow with the continued broadband deployment.

The study suggests that unlike the superpeer architectures, the mutual index caching approach effectively distributes the superpeer functionality over all network hosts, which makes it not only better-performing, but also easier to develop, deploy, manage and maintain. The absence of the explicitly assigned superpeer roles also eliminates the obvious performance bottlenecks and attack targets in such a fully decentralized peer-to-peer network.