Decentralized algorithms are becoming ever more prevalent in almost all real-world applications that are either data intensive, computation intensive or both. This thesis presents a few decentralized solutions for large-scale (i) data dissemination, (ii) graph partitioning, and (iii) data disambiguation. All these solutions are based on gossip, a light weight peer-to-peer data exchange protocol, and thus, appropriate for execution in a distributed environment.
For efficient data dissemination, we make use of the publish/subscribe communication model and provide two distributed solutions, one for topicbased and one for content-based subscriptions, named Vitis and Vinifera respectively. These systems propagate large quantities of data to interested users with a relatively low overhead. Without any central coordinator and only with the use of gossip, we build a novel topology that enables efficient routing in an unstructured overlay. We construct a hybrid system by injecting structure into an otherwise unstructured network. The resulting structure resembles a navigable small-world network that spans along clusters of nodes that have similar subscriptions. The properties of such an overlay make it an ideal platform for efficient data dissemination in large-scale systems. Our solutions significantly outperforms their counterparts on various subscription and churn scenarios, from both synthetic models and real-world traces.
We then investigate how gossiping protocols can be used, not for overlay construction, but for operating on fixed overlay topologies, which resemble graphs. In particular we study the NP-Complete problem of graph partitioning and present a distributed partitioning solution for very large graphs. This solution, called Ja-be-Ja, is based on local search and does not require access to the entire graph simultaneously. It is, therefore, appropriate for graphs that can not even fit into the memory of a single computer. Once again gossip-based algorithms prove efficient as they enable implementing light-weight peer sampling services, which supply graph nodes with partial knowledge about other nodes in the graph. The performance of our partitioning algorithm is comparable to centralized graph partitioning algorithms, and yet it is scalable and can be executed on several machines in parallel or even in a completely distributed peer-to-peer overlay. It can be used for both edge-cut and vertex-cut partitioning of graphs and can produce partition sizes of any given distribution.
We further extend the use of gossiping protocols to find natural clusters in a graph instead of producing a given number of partitions. This problem, known as graph community detection, has extensive application in various fields and communities. We take the use of our community detection algorithm to the realm of linguistics and address a well-known problem of data disambiguation. In particular, we provide a parallel community detection algorithm for cross-document coreference problem. We operate on graphs that we construct by representing documents’ keywords as nodes and the co-location of those keywords in a document as edges. We then exploit the particular nature of such graphs, which is coreferent words are topologically clustered, and thus, can be efficiently discovered by our community detection algorithm.
Stockholm: KTH Royal Institute of Technology, 2014. , x, 22 p.
2014-05-22, Sal/Hall D, KTH - ICT, Isafjordsgatan 39, Kista, 13:00 (English)