An Efficient Search Scheme Using Self-organizing Hierarchical Ring in Unstructured Peer-to-Peer Systems

We propose an efficient search scheme based on a self-organizing hierarchical ring structure in unstructured peer-to-peer systems. Our solution maintains a consistent super-peer ratio and makes the peers with relatively high capacities super-peers in dynamic environments. The benchmarking results show that the proposed algorithm outperforms the static algorithm that has the fixed number of rendezvous peers such as JXTA. In hierarchical unstructured peer-to-peer networks utilizing peers’ heterogeneity, super-peers with larger capacity form a super-peer network and serve their neighbors by storing the peers’ advertisements and forwarding the queries from their neighbors through flooding, random walk, or routing using DHT. While using DHT in rendezvous network makes the search in JXTA faster, there is no control mechanism of the number of super-peers when the network size is changed or the peers’ capacities are skewed in JXTA. Our solution provides an efficient P2P search algorithm based on a self-organizing hierarchical ring that maintains a consistent super-peer ratio in dynamic environment and keeps the peers with relatively high capacities as super-peers. This consists of four algorithms, agent population and distribution control algorithm, peer selection algorithm, query routing algorithm, and the Gossip algorithm. First, all peers collect the capacity information of neighboring peers via mobile agents. In order to reduce the excessive overhead by agents and collect sufficient information under dynamic peer-to-peer network, we should control the agent population appropriately. When an agent arrives at a peer, it computes the elapsed time, t, since the last agent arrived at the peer. If t is less than a termination threshold , the agent terminates itself. But if t is greater than a cloning threshold , a new agent is generated with probability and migrated to another peer as well as the previous one. The probabilistic cloning decision prohibit the excessive agent generation. If t is between and , the agent is migrated another peer based on the migration strategy. Under our migration strategy, each agent is likely to select the neighbor peers with the larger inter-arrival time of agents and smaller number of degree as a migration destination. Second, based on the capacity information of other peers collected by agents, each peer estimates the average capacity of the whole system. The peer whose capacity is ρ * This work was supported by the Brain Korea 21 project in 2006. 56 S. Han, J. Sohn, and S. Park times greater than the average increases its own counting variable, where ρ is determined by the ratio of super-peers. Otherwise, the peer decreases its counting variable. Using two thresholds for promotion and demotion, each super-peer decides to demote or not, and each leaf-peer decides to promote or not at every iteration. Third, unlike other hierarchical systems where peers directly reach super-peers, in order for peers to publish advertisements or send query requests through super-peers, each peer first have to look for any super-peer via random walk. When a super-peer is found, it directs the request to a destination super-peer using hash table. But if the super-peer that should store the requested advertisement does not store the advertisement, it forwards the request to its super-peer neighbors next to it on the super-peer view until the TTL is expired or find the super-peer that stores the advertisement. We call this process the super-peer walk. Last but not least, the super-peers in our system utilize the gossip algorithm [1] as a membership protocol to maintain loosely-consistent super-peer view. When the number of super-peers attending the gossip is N, at least log(N) + c of super peers will be fanout. Therefore, each super-peer generates gossip messages for the log(N) + c peers among the peer view, and then sends them to log(N) + c of randomly chosen super-peers. On receiving gossip messages, the super-peers update their peer view with the peer view information inside the messages. Using an event-driven simulator, we compared the performance of our selforganizing ring algorithm with that of JXTA when the size of the system is gradually increased. As we can see from Fig.1 (a), the number of success messages of our approach is larger than that of JXTA. This indicates that the peer view synchronizing protocol manages to handle the churn situation of super-peers in our approach. Moreover, Fig.1 (b) shows that the average search time of success messages in our approach is better than that in the JXTA. The search time depends largely on not only the traverse hop count but also the peers’ capacity. The self-organizing super-peer ring algorithm allows the peers with higher capacity to be promoted as super-peers and causes the peers with lower capacity to be demoted as non-super-peers continuously. As a result of that, our approach can handle more messages faster than the JXTA system with fixed member of super peers. For the overhead, our approach has more overhead messages than JXTA system due to the additional messages related to agents and the membership protocols as shown in Fig. 1 (c). Fig. 1. (a) Number of Success Messages (b) Average Search Time (c) Overhead