This paper studies a simple strategy, proposed independently by Gallager [l] and Katevenis [2], for fairly allocating link capacity in a point-to-point packet network with virtual circuit routing. Each link offers its packet transmission slots to its user sessions by polling them in round-robin order. In addition, window flow control is used to prevent excessive packet queues at the network nodes. As the window size increases, the session throughput rates are shown to approach limits that are perfectly fair in the max-min sense. That is, the smallest session rate in the network is as large as possible and, subject to that constraint, the second-smallest session rate is as large as possible, etc. If each session has periodic input (perhaps with jitter) or has such heavy demand that packets are always waiting to enter the network, then a finite window size suffices to produce perfectly fair throughput rates. The round-robin method is considerably simpler than earlier strategies for achieving global fairness. The fair session rates are not explicitly computed, and the only overhead communication is that required for the window acknowledgments. The main drawback is that large windows are needed to achieve even approximately fair throughputs in some (hopefully rare) situations, and large windows permit large crossnetwork delays. Fortunately, the round-robin method offers other throughput guarantees that, while falling short of perfect fairness, do apply even for sessions with small windows. Such sessions are promised reasonable bounds on their crossnetwork packet delay as well.

Round robin link scheduling, in conjudction with conventional window flow control, can be used to achieve throughput fairness in point-to-point packet networks with virtual circuit routing. 1. INTRODUCTION Consider a data communication network consisting of store-and-forward nodes joined by point-to-point links. Each user session is assigned a fixed path (often called a virtual circuit) through the network, and data for the session are sent in packets along this path. In such a network it is possible for the incoming traffic rate at a node to exceed the outgoing rate, causing a data queue to build up at that node. This queue may eventually overflow the node's storage 1i This research was conducted at the M.I.T.

In shared-memory packet switches, buffer management schemes can improve overall loss performance, as well as fairness, by regulating the sharing of memory among the different output port queues. Of the conventional schemes, static threshold (ST) is simple but does not adapt to changing traffic conditions, while pushout (PO) is highly adaptive but difficult to implement. We propose a novel scheme called dynamic threshold (DT) that combines the simplicity of ST and the adaptivity of PO. The key idea is that the maximum permissible length, for any individual queue at any instant of time, is proportional to the unused buffering in the switch. A queue whose length equals or exceeds the current threshold value may accept no more arrivals. An analysis of the DT algorithm shows that a small amount of buffer space is (intentionally) left unallocated, and that the remaining buffer space becomes equally distributed among the active output queues. We use computer simulation to compare the loss performance of DT, ST, and PO. DT control is shown to be more robust to uncertainties and changes in traffic conditions than ST control.Index Terms-Adaptive thresholds, asynchronous transfer mode, buffer allocation, dynamic thresholds, memory management, pushout, queue length thresholds, shared-memory switch.

No abstract

Resource reservation must operate in an efficient and scalable fashion, to accommodate the rapid growth of the Internet. In this paper, we describe a distributed architecture for inter-domain aggregated resource reservation for unicast traffic. We also present an associated protocol, called the Border .) BGRP maintains these aggregated reservations using "soft state." To further reduce the protocol message traffic, routers may reserve bandwidth beyond the current load, so that some sources can join or leave the tree without sending messages all the way to the tree root. BGRP relies on Differentiated Services for data forwarding, hence the number of packet classifier entries is extremely small.