Existing protocols, like Mobile IP, provide an initial solution for the mobile host routing problem. However, Mobile IP was not designed to support continuous media streams nor to provide low latency handoffs. The protocols still need to be extended to exploit user tracking and local geographic information, to better support low latency handoffs within and between networks with exposed routing control, and must be better integrated with the emerging protocols for supporting quality of service concepts for real-time and near real-time data streams.
Tracking users and managing their handoffs in regions of high user density generates considerable additional processing load. The key to maintaining low latency in the network management algorithms is to design these algorithms to be able to exploit scalable processing techniques, such as harnessing network of workstations (NOWs). This layer of the architecture will be designed to take full advantage of the distributed processing power of NOWs.
Existing handoff algorithms have been designed for homogeneous wireless subnetworks. Algorithms to support handoff between heterogeneous subnets are considerably more complex. Such algorithms will need to base the handoff decision on application-specified policies as well as periodic measurements of the quality of the underlying network connectivity by the mobile host.
These policies can constrain the feasible handoffs. For example, an application might specify that high priority traffic traverse the available connections with the lowest latency, regardless of cost. Less critical traffic might be constrained to travel over the lowest cost connections currently available. Clearly the goal is to get the data through, so issues like signal quality, bit error rates, and the resulting probability of packet loss and retransmission must be considered as part of the handoff decision process. Under certain conditions, the handoff algorithm might defer switching its connection to a higher latency channel, even though it has better signal quality due to stronger signal strength, from a low latency channel with a weak signal.
Maintaining multiple powered-on network interfaces is an expensive proposition for the mobile host. For reasons of extending battery life, we assume that only one network interface is powered on and selected for providing connectivity to one overlay network at any point in time. Other interfaces will be in a standby mode, and powered up from time to time to determine the quality of connectivity to their respective overlays. A trade-off exists between the frequency of determining link quality, the power consumed by the operation, and how up-to-date the information is about the network state to drive handoff decision making. When the mobile is low on power, a good strategy is to connect to the network with the lowest transmitter power demands, and the handoff algorithm in the mobile may scale back the frequency of probing other overlays. System context may also dictate the frequency of probes. For example, as the mobile moves towards the fringe cells of its current network overlay, a good strategy would be to increase the probes to alternative overlays.
We are experimenting with rules-based languages for describing network selection policy. Such a specification must identify the application's relative priority of characteristics to drive the choice of a particular network for connectivity: highest raw bandwidth, lowest raw latency, lowest usage cost, best overall reliability of delivery (i.e., lowest packet loss), lightest network load, etc.
Within the network management layer, we must also extend existing protocols to improve reliable transport over wireless links and exploit connection-oriented strategies to provide better support for real-time and near real-time media streams. Existing real-time protocols achieve their performance guarantees through admission-control procedures. Once a link becomes fully subscribed, no new connections are allowed to use that link. New schemes for link sharing are being developed to better utilize link bandwidth. Yet even in these schemes, once a connection is granted to an application, it does not need to tailor its demand. These approaches are of limited use in wireless overlay networks, where applications must be able to adapt dynamically to dramatic changes in the quality of their connectivity as they move between networks. Since entire subnets are "black pipes," we must depend on end host adaptation to changes in the network characteristics, rather than strong support from the network, such as resource reservations. And we must extend support for adaptation beyond the applications-independent protocol stack to the actual applications running on the mobile host. We are developing stronger coupling between the underlying networks' capabilities and application demands. This may result in new quality of service protocols, methods for characterizing network quality and availability of alternative connectivity, call-back service alert mechanisms to the application, or combinations of these.