Network coded wireless architecture

Abstract

Wireless mesh networks promise cheap Internet access, easy deployment, and extended range. In their current form, however, these networks suffer from both limited throughput and low reliability; hence they cannot meet the demands of applications such as file sharing, high definition video, and gaming. Motivated by these problems, we explore an alternative design that addresses these challenges. This dissertation presents a network coded architecture that significantly improves throughput and reliability. It makes a simple yet fundamental switch in network design: instead of routers just storing and forwarding received packets, they mix (or code) packets' content before forwarding. We show through practical systems how routers can exploit this new functionality to harness the intrinsic characteristics of the wireless medium to improve performance. We develop three systems; each reveals a different benefit of our network coded design. COPE observes that wireless broadcast naturally creates an overlap in packets received across routers, and develops a new network coding algorithm to exploit this overlap to deliver the same data in fewer transmissions, thereby improving throughput. ANC pushes network coding to the signal level, showing how to exploit strategic interference to correctly deliver data from concurrent senders, further increasing throughput. Finally, MIXIT presents a symbol-level network code that exploits wireless spatial diversity, forwarding correct symbols even if they are contained in corrupted packets to provide high throughput reliable transfers. The contributions of this dissertation are multifold. First, it builds a strong connection between the theory of network coding and wireless system design. Specifically, the systems presented in this dissertation were the first to show that network coding can be cleanly integrated into the wireless network stack to deliver practical and measurable gains. The work also presents novel algorithms that enrich the theory of network coding, extending it to operate over multiple unicast flows, analog signals, and soft-information.

(cont.) Second, we present prototype implementations and testbed evaluations of our systems. Our results show that network coding delivers large performance gains ranging from a few percent to several-fold depending on the traffic mix and the topology. Finally, this work makes a clear departure from conventional network design. Research in wireless networks has largely proceeded in isolation, with the electrical engineers focusing on the physical and lower layers, while the computer scientists worked up from the network layer, with the packet being the only interface. This dissertation pokes a hole in this contract, disposing of artificial abstractions such as indivisible packets and point-to-point links in favor of a more natural abstraction that allows the network and the lower layers to collaborate on the common objectives of improving throughput and reliability using network coding as the building block. At the same time, the design maintains desirable properties such as being distributed, low-complexity, implementable, and integrable with the rest of the network stack.