Congestion control in networks with no congestion drops

Congestion is intrinsic to the operation of networks and is usually handled by a combination of algorithms at the link and network/transport layers. Link level algorithms alleviate "transient congestion" caused by the temporary oversubscription of a link due to a burst of packets arriving at a switch or router buffer. Network or transport level algorithms alleviate "sustained congestion" which occurs when the long- term arrival rate at a link exceeds its capacity. Algorithms at the two levels interact to provide a scalable, stable and fair bandwidth allocation to the flows passing through the network. Link level algorithms are typically very simple: drop or mark packets with increasing probability as buffer congestion increases; moreover, if a packet arrives at a full buffer, drop it. These dropped or marked packets are used by the transport algorithms to adjust the traiisniissioii rate of sources. In this paper we are concerned with networks in which packets cannot be dropped when there is congestion. In such networks a back-pressure mechanism "pauses" the link or links feeding a congested buffer, thus preventing further packets from arriving at the buffer. The links are later unpaused when the buffer becomes uncongested. This paper is a theoretical study of the stability and fairness properties of network level congestion control when pause mechanisms operate at the link level to prevent packet drops. Our focus is on the Backward Congestion Notification (BCN) algorithm which is being considered by the IEEE 802.1 standards body for deployment in switched Ethernet networks.