Wired networks are traditionally viewed as a unicast medium: packets are sent from A to B. Because the underlying medium can have a degree of sharing (e.g. multiple hosts connected to the same Ethernet hub), steps must be taken to avoid interference between concurrent senders (e.g. CSMA).
In a wireless network, the medium is shared: any hosts in range of a radio transmission can receive it. If the goal is to use a traditional unicast routing protocol over a wireless link, this makes interference between senders a more challenging problem (e.g. as discussed in the MACAW paper). However, rather than viewing broadcast as an inconvenience we need to work around, the broadcast nature of wireless can also be leveraged to design new protocols that would be infeasible for a wired network.
A nice example of a broadcast-oriented protocol for wireless is “ExOr: Opportunistic Multi-Hop Routing for Wireless Networks“. In a traditional routing design, a sender chooses the next hop that should receive a packet, and then unicasts the packet to that destination. In ExOr, a sender broadcasts a batch of packets to a group of nodes simultaneously. The set of batch recipients coordinate to try to ensure that each packet in the batch is forwarded onward. The recipient of a packet is only chosen after the packet is sent, which allows ExOr to “opportunistically” take advantage of links that have high loss rates. ExOr assumes that node reception probabilities are mostly independent and mostly decrease with distance — both of which are probably pretty reasonable assumptions.
Design
The source collects a batch of packets destined for the same host, and then chooses a forwarding list for the batch. The forwarding list is a list of nodes, sorted by the expected cost of delivering packets from that node to the eventual destination node. The cost metric is similar to ETX (expected number of transmissions required to send the packet to the destination via unicast, including retransmissions); unlike ETX, it only considers the forward delivery probability. While it would be possible to including all possible recipient nodes in the forwarding list, this would increase the coordination cost among the forwarders, so ExOr only includes “likely” recipients in the list (estimated 10%+ chance of receiving a broadcast packet).
Each packet contains a batch map, which holds the sender’s estimate of the highest priority (according to the cost metric) node to have received each packet in the batch. When a node receives a packet, it uses the packet’s batch map to update its local batch map. This means that batch map information propagates through the nodes, carrying information about packet reception from high priority nodes to lower priority nodes.
After the source broadcasts a batch, each member of the forwarding list broadcasts, ordered by descending priority (ETX value). Each node broadcasts the packets it received, along with its updated batch map. Nodes coordinate to schedule their transmissions to try to avoid interference (e.g. by estimating when each node’s expected transmission time is, according to ETX value and batch map contents). The protocol continues cycling through the nodes in priority order. At the end of each cycle, the ultimate destination broadcasts its batch map 10 times; at the beginning of each cycle, the source resends packets that weren’t received by any node (by observing batch map contents). The ExOr scheme stops when 90% of the packets in a batch have been transferred, and uses a traditional routing policy to deliver the remainder of the batch.
Discussion
Overall, ExOr is a really neat idea. That said, it is at best a special-purpose solution, because of the high latency it incurs for batching and multiple transmission rounds. Similarly, the need for a split TCP proxy is pretty ugly. A complete solution to wireless routing would perhaps adaptively switch between latency-oriented and bandwidth-oriented routing techniques, depending on the nature of the traffic.
It seems arbitrary to me that ExOr works on a batch-by-batch basis—that almost seems like establishing a new TCP connection for each window’s worth of data. The amount of useful work done in each transmission round decreases, until the protocol imposes an arbitrary cutoff at 90% and switches to a traditional routing protocol. Instead, wouldn’t it be more sensible for ExOr to allow new packets to be inserted into the batch as the destination confirms the delivery of packets? This would essential emulate the “conservation of packets” principle from the Van Jacobson paper.