Research On Highly Robust Data Distribution Technology For Dynamic Network Environment

With the development of networking technology and the growth of demand forsharing information, many distributed applications based on data distribution aredeveloped. The common demand of these applications is to distribute various datagenerated dynamically by sources to users with different interests rapidly and accuratelyin dynamic network environments. Therefore, two features of the data distributionapproach—the robustness against the dynamic network environment and the abilityof distributing data efficiently—are simultaneously demanded. As for the applicationssuch as emergency management, network-centric warfare and decentralized socialnetworking service, higher robustness is especially needed due to the dynamicity oftheir network environments. According to the different ways the users expressing theirinterests, the existing data distribution approaches can be approximately divided intothree types: content-based approaches, topic-based approaches and approaches oversocial networks. This dissertation deeply studies the techniques for content-based datadistribution, topic-based data distribution and data distribution over social networks inorder to realize highly robust data distribution in the dynamic network environment.Since the existing content-based data distribution approaches mainly lack ofbalance between efficiency and robustness, this dissertation proposes a content-baseddata distribution approach over self-organizing semantic overlay networks—SemanticCast—to realize the efficient content-based data distribution in the dynamicnetwork environment. SemanticCast maintains a self-organizing semantic overlaynetwork based on view exchange, called Crowd, in which, each node tries to retain theneighbors with more similar interests during the periodical view exchange. Through thisself-organizing behavior of nodes, various interest clusters, which are constituted by thenodes with similar interests and have no clear boundaries, emerge in the overlaynetwork. Over Crowd, SemanticCast adopts random walk to route data between interestclusters, and adopts flooding to disseminate data inside the interested cluster. Theexperimental results reveal that compared with existing approaches, SemanticCastrealizes more efficient content-based data distribution in the unreliable and dynamicnetwork environment. What’s more, the strong self-healing ability of SemanticCast canmake it recover from the major network disaster quickly.Since the existing topic-based data distribution approaches lack of balance betweenefficiency and robustness, this dissertation proposes a topic-based data distributionapproach over hybrid overlay networks—Laurel—to realize the efficienttopic-based data distribution in the dynamic network environment. Laurel organizesnodes into clusters according to their interests to reduce the unnecessary data dissemination. It adopts the structured inter-cluster topology to route data efficiently,and adopts multiple connections between clusters and unstructured topologies insideclusters to ensure robustness. Laurel firstly routes data to the interested cluster by thestructured inter-cluster topology, and then disseminates data inside the interested clusterby flooding or gossip. The experimental results reveal that compared with theapproaches based on the structured inter-cluster topology, Laurel ensures evidentlyhigher robustness along with the same inter-cluster routing efficiency, and comparedwith the approaches based on the unstructured inter-and intra-cluster topology, Laurelensures evidently higher efficiency for inter-cluster routing along with the comparativerobustness. At the same time, better load balance within clusters is achieved.Under the circumstance of a single node having relatively large numbers of topics,the existing topic-based data distribution approaches cost a lot and fail to work well.Therefore, this dissertation proposes a topic-based data distribution approach based ontopic sampling—TopicCast—to realize the scalable topic-based data distribution inthe dynamic network environment. TopicCast consists of two relatively independentcomponents: a gossip-based topic sampling approach TopicSampler and a lightweighttopic-connected overlay network protocol TopicGraph. TopicSampler firstly estimatesthe proportions of nodes with various topics through the peer sampling service based ongossip, and then uses the estimated proportions to maintain each node’s topic samplingtable, where the nodes with different topics occur randomly under the same probability.Meanwhile, TopicGraph updates each node’s neighbor list periodically by using nodesamples provided by the peer sampling service, and tries to ensure the connectivity ofeach subgraph induced by the nodes with the same topic (called topic-connected) at alow cost. TopicCast firstly directs data to any interested node by the topic samplingtable, and then disseminates data to all interested nodes over the matchingtopic-connected subgraph. Both of the theoretical and experimental results reveal thatTopicSampler can achieve precise proportion estimation and approximately uniformrandom topic sampling in the dynamic network environment, and TopicGraph canensure topic-connectivity of the overlay network with low storage and communicationcost in the dynamic network environment, which means that TopicCast not only realizesthe efficient topic-based data distribution in the dynamic network environment but alsohas better scalability compared with existing approaches.The existing decentralized data distribution approaches over social networks eitherlack of robustness against the highly dynamic network environment or the ability toimplement the data distribution function provided by centralized social networkingservices. This dissertation proposes a P2P data distribution approach over socialnetworks—PeerChatter—to realize the robust and efficient decentralized datadistribution over social networks. PeerChatter maintains an overlay network based on multi-level random graphs, called SkipCluster. The nodes represent users in the socialnetwork are organized into SkipCluster. The regular relation between levels enablesSkipCluster to support efficient routing, and the randomness inside the levels ensureshigh robustness against node churn. Over SkipCluster, PeerChatter adopts thetopic-based publish/subscribe model to realize the synchronous and asynchronousmulticast for data distribution between friends and inside groups. Both of the theoreticaland experimental results reveal that PeerChatter’s routing performance achieves andeven exceeds the one of the typical structured topology, and PeerChatter ensures highreliability and efficiency of data distribution in the network environment with a veryhigh node churn or a major network disaster, which means that the demand of thedecentralized data distribution over social networks can be satisfied.