Dual queue approach to improving network performance during transient congestion episodes

Despite extensive work on the network congestion[Jain, 1990], network congestion remains a real problem. The solution is not one answer, but many. Call admission controls try to limit the number of sessions according to the network's traffic capacity. Implicit and explicit feedback congestion control techniques are often too slow in reacting to transient conditions. Window, automatic code gapping, and percentage throttling based techniques all exhibit these shortcomings. Explicit feed back techniques may even make the congestion worse. All these techniques play an important part in reducing congestion. However, transitory periods of congestion still occur. Reasons for this include the time taken to react to congestion, the nature of congestion control, and the character of packet traffic itself [Leland et. al., 1994]. For real time services, such as voice, this causes severe degradation in quality or complete loss of service. For interactive services, such as WWW browsing, delays become intolerable and packet retransmission further aggravates the problem. Unfortunately, such problems are likely to occur at relatively low average utilisation levels due to the self-similar nature of most data network traffic [Erramilli et al., 1996]. Given that transient congestion episodes are still a problem, it is necessary to make the network as tolerant as possible to these. This paper proposes an alternative philosophical approach to the problem and demonstrates that it can be implemented using a system of two queues.