Architecture and Evaluation of an Unplanned 802.11b Mesh Network; Bickets, Aguayo, Biswas, & Morris

There are 2 families of community wireless networks: well-planned ones with good throughput and connectivity, and unplanned ones with minimal setup required but with worse coverage. This paper proposes the architecture for a wireless network that combines the best of both worlds, namely a network that's quick and easy to setup but that still gets good performance; this architecture consists of unconstrained nodes, omni directional antennae, and link-quality-aware multi-hop routing optimized for throughput. It examines the feasibility of this architecture as implemented in Roofnet, consisting of 37 roof-mounted antennae nodes over 1.5 square miles of Cambridge, MA.

Each node has a non-globally routable IP address and assigns IP addresses via DHCP to hosts attached to its Ethernet ports. Hosts can communicate with one another via their nodes due to the NAT between the Ethernet and Roofnet at each node. There are 4 nodes serving as gateways and acting as NATs from Roofnet to the Internet.

Routes are chosen through Roofnet by Srcr, which chooses based on lowest estimated transmission time (ETT), which presumably corresponds to highest throughput. Srcr tends to favor more short fast hops over longer-distance hops. Additionally, SampleRate selects the best bit-rate to send data at over each link.

The authors evaluate the throughput of Roofnet based on all pairs of nodes within the network. The average throughput was around 600 kbits/sec, with throughput being very closely tied to number of hops between the two. Interestingly, 10% of node pairings failed to find routes between them due to link losses that weren't overcomable by Srcr's flooded queries repeated after 5 seconds. Use of omni-directional antennae appears to be useful because most nodes seem to vary their first hop. While Roofnet is fairly robust to loss of links, removing only two of the best-connected nodes cuts throughput in half. However, in evaluating Roofnet, the authors did not affect Internet access of Roofnet users, meaning that non-experiment traffic could have skewed the findings.