Active Queue Management With Variable Bottleneck Rate

The combination of the following trends in packet-switched networking is poised to revive major interest in active queue management (AQM) schemes for controlling the allocation of network buffers to TCP packets: the persistence of TCP as the vastly dominant type of traffic crossing the Internet, the demand for ever-increasing link rates, and the economic pressure for energy-efficient system designs. By enabling much smaller buffer sizes than those required by the conventional tail-drop policy, AQM damps the negative effects of packet memories on link capacity, power consumption, and system scalability. While the performance of a broad variety of AQM schemes has been well characterized under stationary link conditions, very little is known about their behavior with dynamic bottleneck rates. We set forth to start filling this critical knowledge gap. We first recognize that AQM has generally very limited power against a range of bottleneck rate functions and provide simple guidelines for identifying the boundaries of that range. Then we investigate the dynamic properties of Periodic Early Detection (PED), an AQM scheme of recent invention with excellent steady-state performance, and devise enhancements of its specification that substantially improve its response to bottleneck rate variations.