Probabilistic modeling in support of preemptive congestion control

Congestion control schemes seek to perform a number of functions on the Internet. Beyond the fair sharing of network resources, one of these functions is to increase average link utilization by determining the dynamic operating characteristics of network links and tailoring a congestion event response to those characteristics. One of the drawbacks of current congestion control models is their a posteriori operation based on observed measurements. This paper describes a framework for a high resolution, probabilistic model capable of short-term admission control and transmission plan production for both TCP and UDP traffic flows. A transmission plan has the opportunity of operating in a future time frame, rather than in the reactive mode of current systems. Probabilistic modeling also introduces the possibility of smoothing characteristically bursty traffic using distributed, globally optimum control. Since bursty traffic on a heavily loaded link can lead to unwanted phase effects, reduced throughput, and recurrent drops, being able to smooth these traffic patterns could effectively benefit both link utilization and network stability.