Distributed Explicit Rate Allocation

Abstract

The problem of congestion niaiiagement has been a particularly active area of research for the last decade. Networks are getting faster. arc more richly interconnected and are growing in scale everyday. The larger bandwidthdelay products for these high speed networks poses a continuing challenge for flow and congestion management algorithms. Traditional work i n congestion control and avoidance algorithms has been to layer them on top of window flow control. which is an end-to-end mechanism to assure no loss in the receiving end-system. When used in the context of the large bandwidth-delay product seen in high speed networks, window based flow control mechanisms result in the use of large windows and the potential for large bursts of packets arriving at intermediate nodes. This requires significant additional buffering that is often considered undesirable. One of the changes for congestion management in high speed networks is the use of rate based congestion/flow control schemes. The demarkation between flow and congestion control also tends to be blurred with the use of rate based schemes. However, using rate based mechanisms results in a fresh set of problems to solve in the congestion control and avojdance domain. Another fundamental change over the same time frame is the development of switched networks, exemplified by ATM. The issue of dealing with congestion in connection oriented networks, while apparently easier, is complicated by the need to address qualities of service. A lot of the recent work in congestion control has been in the context of ATM flow and congestion control. M u c h of the work has been in the context of the ABR (Available Bit Rate) service. These typically arc used for data transfer, and are considered to be lower priority than the other classes of service, CHR and VIjH. Left over bandwidth from the use of a resource (link, switch or end-system) by CBR and VBR is used for AHR service. The timely use of the available resources by ABR without causing congestion (which results either in significantly increased queueing delays or cell loss) is an important goal. Another equally important goal is to ensure that the resources in the network are shared in a ‘fair and equitable’ manner. A definition of ‘fairness‘ is the notion of max-min fairness, which has been adopted in the ATM Forum for congestion management of the ABIi service. A simplistic notion of inax-min fairness viewed by a flow at a local resource is: If my flow is limited at this resource for lack of adequate capacity. then there should be no other flow passing through this resource that will be given a greater share of the resource’s capacity. There are several ways of computing the max-min fair allocation for a flow in a network of multiple resources and multiple flows with different resource demands [l], [ 7 ] , [8], [12],