A theoretical analysis of feedback flow control

Congestion is a longstanding problem in datagram networks. One congestion avoidance technique is feedback flow control, in which sources adjust their transmission rate in response to congestion signals sent (implicitly or explicitly) by network gateways. The goal is to design flow control algorithms which provide time-scale invariant, fair, stable, and robust performance. In this paper we introduce a simple model of feedback flow control, in which sources make synchronous rate adjustments based on the congestion signals and other local information, and apply it to a network of Poisson sources and exponential servers. We investigate two different styles of feedback, aggregate and individual, and two different gateway service disciplines, FIFO and Fair Share. The purpose of this paper is to identify, in the context of our simple model, which flow control design choices allow us to achieve our performance goals . Aggregate feedback flow control, in which congestion signals reflect only the aggregate congestion at the gateways, can provide time-scale invariant and stable performance, but not fair or robust performance. The properties of individual feedback flow control, in which the congestion signals reflect the congestion caused by the individual source, depend on the service discipline used in the gateways. Individual feedback with FIFO gateways can provide time-scale invariant, fair, and stable performance, but not robust performance. Individual feedback with Fair Share gateways can achieve all four performance goals. Furthermore, its stability properties are superior to those of the other two design choices. By making robust and more stable performance possible, gateway service disciplines play a crucial role in realizing effective flow control .