← Back to the chapter · Previous detail · Next detail →
Technical detail · Edge, Fog, and Cloud: Operating at Scale
Radio selection: stability and efficiency
9.7.3 Decentralised RAT Selection: Stability Is Not Efficiency
A device with more than one radio access technology (RAT) can choose the path that appears to offer the best useful throughput. The choice is coupled to everyone else’s choice: moving to a Wi-Fi access point or cellular cell changes its load, which changes the payoff seen by the next device. One useful model treats each device as a player, the available RATs as strategies, and predicted throughput as the payoff. A stable assignment is a Nash equilibrium when no single device expects to improve by switching alone.
The throughput model must match the shared medium. In one class of network, a user’s result depends on the rates of all attached users; a slow participant can change airtime consumption for the group. In another, capacity is divided mainly by the number of attached users. When every candidate RAT follows the same class, repeated improvement choices can converge under suitable assumptions. Mixing the classes is harder: one user’s apparently beneficial move can change another user’s incentive and create an endless improvement path.
Hysteresis makes the client less reactive. Require the predicted improvement to exceed a margin for a persistence window, optionally randomise the switch among eligible clients, and retain a larger margin when crossing between media with different sharing rules. This reduces ping-pong handovers, but a larger margin also creates more stable assignments that may be farther from the system-wide best result. Faster convergence, fewer switches, battery cost, fairness, and total throughput therefore belong in the same parameter study.
Equilibrium is not proof of a good allocation. Compare the decentralised outcome with a feasible system optimum and report the efficiency loss, the worst-served group, and the distribution of outcomes rather than only an average. A controller can be stable while leaving capacity unused or concentrating poor service on devices whose alternative RAT is weak. The result also depends on assumptions: throughput estimates may be noisy, radios may interfere, network assistance may exist, and a device may use more than one interface. Test inference error, stale load observations, join/leave events, and re-convergence instead of treating perfect measurements as a deployment property.