The Control of Congestion in , Packet - Switching Networks

Any communication network has a finite traffic capacity and if it is offered traffic beyond the limit it must reject some of it. The data-communication network studied here is one employing packet switching, like the Advanced Research Project Agency (ARPA) network. It handles blocks of data, called packets, and longer messages are subdivided, rather in the same way that store in a computer is allocated in pages. A method of controlling congestion is proposed in which there is a finite number of packet’carriers in the whole network. When a packet of data is delivered to its destination node the ccempty” packet is available for reuse. The empties move randomly round the network and new data must capture an empty packet carrier before being launched into the network. Various elaborations are described that avoid delay in normal conditions. This so-called Manuscript received May 10, 1971; revised January 10, 1972. This work was performed in part by Plessey Telecommunication Research, Taplow, Bucks., England. This paper was presented at the 2nd Symposium on Problems in the Optimization of Data Communications Systems, Palo Alto, Calif., October 20-22, 1971. The author is with the National Physical Laboratory, Teddington, Middlesex, England. Ysarithmic” method of congestion control supplements and does not replace end-to-end flow control. CONGESTION IN DATA-COMMUNICATION NETWORKS A NY communication network has a limit to the traffic it can carry. If there is more than a certain traffic demand, some of the traffic must be rejected. Both the nature of the limitation and the reaction of the network to excess demand depend on the design of the network. The network in a condition where it must reject traffic is called “congested.” The avoidance of congestion is of great importance to public data networks because the facilities they offer will be built in to computer systems that are vital to the operation of trade, industry, transport, etc. Good planning and provision to meet demand is the only means of avoiding congestion. The safety margin needed to guard against congestion will therefore be high for data networks. DAVIES: CONTROL OF CONGESTION In the same way that strenuous efforts are made to avoid faults in computer systems (but the reaction of the system to faults is carefully studied and engineered), so i t is necessary to study the reaction of data-communication networks to congestion, even though congestion is to be avoided by adequate provision. Fortunately the approach t o congestion can readily be monitored quite independently of its deleterious effects. Therefore it is not necessary to design systems in which the impairment of performance near congestion is used to give warning of failure. This simple point seems to have been misunderstood by some commentators. We can therefore strive by good design to reduce the effects of overload while planning for the network not to reach this condition and monitoring the safety margin continuously. This paper concerns the control of congestion in a particular kind of data network that employs packet switching. PACKET SWITCHING IN DATA NETWORKS Packet switching is a variant of the message-switching principles used to handle telegraphic messages, but its aims and therefore its design details are different. It is characterized by a low figure of queueing delay, which can be about 10 ms for one transit through a national network using today’s technology. A good current example of the packet-switching method is the Advanced Research Projects Agency (ARPA) network [ 11. The low delay figure is achieved by employing fast links and short message units. Because the blocks of data to be moved by the network are of variable length but predominantly short, a message unit of 1000 bits or less is proposed. Larger blocks to be moved will be broken into these small units, rather in the way that computer storage is allocated in pages of constant length while the user, unaware of this, deals in segments. The same sort of distinction is made here by calling the customer’s unit a “message” and the network unit a “packet.” This term is the origin of the name “packet switching.” Communication through the network is usually a matter of sending packets back and forth between two subscribers. While they communicate, a link is said to exist between these subscribers. But a subscriber may also be a multiaccess computer or a cluster of terminals connected to the network through a local packet switch. Such a subscriber must be able to establish many links a t one time. It is useful to. have a word to describe the individual source or destination of data, whether i t is a simple terminal, a process in a multiaccess computer, or one of a cluster of terminals multiplexed by the subscriber. We call i t a “socket,” using the ARPA terminology. A bidirectional terminal has two sockets, one source, and one destination. Links are therefore stablished between sockets, bidirectional inks employing two sockets a t each end. One purpose of the link concept is to simplify the task of a simple terminal attached to the network. For such 547 a simple terminal, outgoing packets must be assembled and formated by the network and provided with a destination field in their headings. Incoming packets must be rejected unless they come from the link-designated source. Note that the link is a software feature. A succession of packets moving from source to destination forms a data-communication channel that has a variable data rate. Because links can have variable capacity, a technique of data-rate control must be developed for these networks. The ultimate limit to traffic in packet-switching networks is expressed in terms of information carried (e.g., packets/second) whereas the limit applying to circuitswitched networks is measured in numbers of calls. It is possible, in principle, for a packet-switching network to react to congestion by reducing the effective data rate of certain links. EXISTING METHODS FOR CONTROL OF CONGESTION By analogy with road tra.ffic, congestion can be expected to begin a t one point in the network and spread as the queues fill and links between switching centres are blocked. Good control of the route taken by packets can increase the load the network will take, but when the limit is eventually reached, several links or nodes will tend to be blocked simultaneously. Existing congestion-control methods can be classified as local or end to end. Local control is applied by a switching center on the basis of its own local traffic data (packet rates, queue lengths) also using operational messages received from its immediate neighbors. These messages may request a reduction of traffic over a particular link, or restore unrestricted working, or they may contain data on traffic or delays experienced, etc. They come only from neighbors. Eventually, as traffic levels increase, the rerouting of packets can no longer prevent congestion, and the network must reject traffic offered to it. If the users who chiefly contribute to an overload are distant from the point of congestion, local control methods require congestion-control measures to spread a long way before effective measures are taken. This is not to say that local control is necessarily ineffective, but it presents difficult design problems. End-to-end control makes use of the notional links that exist between subscribers (or, more correctly, between sockets). New links that might cause congestion can be refused, in the same way that circuit-switched networks operate. There are some objections to this method. It controls the wrong parameters because a link is not associated with any particular data rate. The variable-rate nature of links must be preserved, otherwise the designer of a teleprocessing system using the network has to look after data rates as well as the logical structure of the necessary links. A valuable feature of packet switching is the ability to hold open a link economically and thus get