A novel congestion control framework for delay and disruption tolerant networks

• Autor(a) principal: Aloizio Pereira da Silva
• Data de Publicação: 2015
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações do ITA
• Texto Completo: http://www.bd.bibl.ita.br/tde_busca/arquivo.php?codArquivo=3405
• topic: Comunicação sem fio; Processamento de sinais; Atraso; Intermitência; Modelos matemáticos; Redes de comunicação; Engenharia eletrônica; Computação

Delay and Disruption Tolerant Networks (DTNs) are networks that experience frequent and long-lived connectivity disruptions. Unlike traditional networks, such as TCP/IP Internet, DTNs are often subject to high latency caused by very long propagation delays (e.g. interplanetary communication) and/or intermittent connectivity. In DTNs there is no guarantee of end-to-end connectivity between source and destination. Such distinct features pose a number of technical challenges in designing core network functions such as routing and congestion control mechanisms. Detecting and dealing with congestion in DTNs is an important and challenging problem. Existing DTN congestion control mechanisms typically try to use some network global information or they are designed to operate in a particular scenario and they depend on forwarding strategy, for example, replication forwarding. As a result, existing DTN congestion control mechanisms do not have good performance when they are applied in different scenarios with different routing protocols. In this thesis, we first review important challenges of DTNs and survey the existing congestion control mechanisms of this domain. Furthermore, we provide a taxonomy of existing DTN congestion control mechanisms and discusses their strengths and weaknesses in the context of their assumptions and applicability in DTN applications. We also present a quantitative analysis of some DTN congestion control mechanisms to evaluate how they behave in deep space communication scenario since they were designed to operate at terrestrial DTN. We extensively evaluated these mechanisms using two different applications and three different routing protocols and mobility patterns. The evaluation results show that the selected mechanisms poorly perform in deep space scenario. Therefore, in view of DTN characteristics, to study new congestion controls and better undersand the impact of congestion in DTN we modeled DTN congestion problem using percolation theory. We formulate the DTN congestion problem as a percolation process resulting in a percolation model that is simple and easy to derive. Another important feature of the proposed percolation model is the fact that instead of requiring global information about the whole network, it relies exclusively on local information, i.e., information related to a node and its neighboring nodes. The principal advantage of our mathematical model is to provide a fast way of having an idea of the system's performance being modeled and allow us to validate either simulation or realistic experiments. Consequently the proposed model can be used to predict and control congestion in DTNs. Being aware that far from the traditional network, DTN is a new kind of network derived by deep space communication and as congestion control is an important factor that directly affects network performance. The development of DTN must rely on the perfect congestion control mechanism to ensure reliability, stability and extensiveness of the network. In order to enhance the reliability of data delivery in such challenging network, this thesis proposes DTN-Learning, an adaptive and autonomous congestion aware framework that mitigates the congestion by using Reinforcement Learning. This allows the network nodes to adapt their behavior on-line in a real environment. It is general and can easily be combined with existing schemes for local control. Preliminary results show that using our adaptive approach, the network node exhibits a more accurate behavior, increasing the delivery ratio and decreasing drop ratio, as compared to approaches that do not use learning. This mitigates congestion phenomena observed in non-adaptive local congestion control mechanisms and helps the network to reach high performance faster.