Distributed control algorithms for adapting publish/subscribe in software defined networks

Abstract

Content-based routing has emerged as a popular paradigm in publish/subscribe systems for interactions between its system components (publishers and subscribers). Content-based routing of published information is extremely bandwidth efficient as a publication is forwarded only to subscribers which have expressed their interest in this published content. Rules or filters are applied on the content at one or more routers/brokers to determine its path through the network. Most of the state of the art solutions consist of a distributed set of brokers scaling wide-area networks. However, in each of these solutions, filtering of events takes place at the application layer at dedicated brokers. This expensive filtering phase renders the advantages of content-based pub/sub with regards to bandwidth efficiency less significant as it results in higher end-to-end latency and lower throughput rates. To overcome this problem, software-defined networking may be used to build publish/subscribe systems where filtering of events can happen directly in the Ternary Content-Addressable Memory (TCAM) of network routers. Initial work has shown that it is possible to map effectively a content-routing topology to network routers controlled by a single broker, resulting in line-rate forwarding of data packets. However, a single broker limits the performance of the system with regards to scalability and is not feasible in large networks consisting of numerous network elements. To incorporate the best of both worlds, this thesis proposes distributed control algorithms using software-defined networking that allow to build a publish/subscribe system spanning over multiple sub-networks of controllers where the controllers divide the network spatially into disjoint partitions. With respect to such an architecture, this thesis discusses the manner in which connectivity is established between sub-networks along with the costs incurred in the process. A detailed analysis of the average controller overhead and total control traffic generated in the proposed system is presented which is further supported by simulation results. It also includes an analysis on the effect of distributing control on certain performance metrics such as false positive rate of published events.