Congestion Control for Best-Effort Service: Why We Need a New Paradigm - IEEE Network

he first generation of telecommunication networks were circuit-switched. In such networks, each circuit is allocated resources along a path for its exclusive use; there is no uncertainty about the bandwidth or delays along the path. Packet switching introduced a very different mode of communication. In most packet-switched networks,l sources send their data packets into the network without any prereserved resources, and the network exerts its “best effort” to service the packets. The advantage of packet switching is that it allows the network resources (bandwidth, buffers) to be statistically shared among all sources. This is especially important for computer communications, since data traffic tends to be rather unpredictable and bursty; prereserving resources would lead to low utilization levels, whereas the statistical multiplexing of packet switching allows one to achieve much higher utilization levels. The quality of service (in terms of packet delivery delays and drops due to buffer overflow) of best-effort service depends not only on the network actions (which the network can control), but also on the offered load (which the network cannot control). Thus, in best-effort service, the network tries to forward all packets as soon as possible, but cannot make any quantitative assurances about the quality of service delivered. The unpredictable and bursty nature of computer traffic not only prevents the network from making quality assurances, but also creates the problem of congestion. The ratio of average demand to peak demand on any particular link is quite high. Thus, while it is advisable to provision the network to have adequate resources to satisfy the average demand, it is not economically feasible to provision the network to satisfy the peak demands. Consequently, there will be times when the network is momentarily overloaded when many sources happen to send their data simultaneously. To control this congestion, the network must provide feedback to the users indicating that the network is currently, or is about to become, overloaded; in response to such a congestion signal, the users should inject their packets into the network more slowly. These processes, the feedback from the network and the source response, form the fundamental basis of congestion