Improving Throughput, Latency and Privacy with Hybrid Networks and Multipath Routing

Over the last few years, residential and enterprise networking have faced several challenges due to the increasing demand of users for high-throughput connectivity. As a result, efforts are being made to improve coverage, throughput, and robustness. Several solutions have been recently proposed. The first solution is to use mesh networking; it is gaining momentum, as it effectively improves performance, but at the cost of an increased complexity compared to the infrastructure mode, as several paths can now be employed with potentially several hops. The second solution is to exploit the different technologies that are available, wired (e.g., power-line communication (PLC) or Ethernet) and wireless (e.g., WiFi or cellular). Networks with various technologies are referred to as hybrid networks. When the technologies do not interfere with each other, it is possible to aggregate their capacity, thus enabling immediate throughput improvements; by increasing the number of possible paths that a packet can take, hybrid networks also increase complexity. The third solution is to use multipath routing, which can improve performance significantly. But again, this comes at the cost of an increased complexity. In this dissertation, we study the effect of these solutions in terms of throughput and coverage, latency, and privacy. We focus, in particular, on hybrid networks with shared-medium and orthogonal technologies, where two links that use the same technology are subject to interference (shared-medium), but not two links that use two distinct technologies (orthogonal). First, we study the effect of these solutions on throughput and coverage. We show that, in hybrid mesh networks, the optimal number of paths achieving maximal throughput with multipath routing is tightly linked with the number of technologies. This result makes it possible to develop an efficient and practical multipath routing protocol that yields the maximal throughput. Next, we introduce two novel algorithms for optimizing throughput: A distributed multipath congestion controller that, when each flow uses one multipath fixed in advance, provably achieves optimal throughput and an algorithm based on the multi-armed-bandit framework that finds the best multipath and converges to the best achievable throughput. We implement these algorithms in a real testbed with PLC and two orthogonal WiFi channels. Their experimental evaluation shows that using technologies with distinct physical layers, such as PLC and WiFi, improves spatial diversity compared to using multi-channel WiFi and brings further improvements of throughput and coverage. Then, we investigate latency in hybrid networks. We study analytically how the variance of a time-varying service rate affects queueing delays. We also study latency when multipath routing is used, i.e., when traffic is split between two technologies. We show that finding the optimal splitting scheme is difficult, as it depends on the rate at which packets arrive, and that the best static scheme, where the splitting probability remains the same for all arrival rates, can be significantly sub-optimal in time-varying networks. Finally, we study how hybrid networks and multipath can improve privacy. We show that they can significantly improve the resistance against traffic analysis attacks, such as website fingerprinting, by enabling the user to split the traffic between two networks.