Wireless Ad Hoc and Sensor Networks

Congestion Control in ATM Networks and the Internet 81Because the ATM forum decided to use a closed-loop rate-based congestioncontrol scheme as the standard for the ABR service (Lakshmanet al. 1999), several feedback control schemes were proposed in the literature(Jagannathan 2002, Chen et al. 1999, Fuhrman et al. 1996, Cheng andChang 1996, Bonomi and Fendick 1995, Benmohamed and Meerkov 1993,Lakshman et al. 1999, Jain 1996, Mascola 1997, Qiu 1997, Izmailov 1995,Bae and Suda 1991, Liew and Yin 1998, Liu and Douligeris 1997, Bonomiet al. 1995, Jagannathan and Talluri 2002) with (Chen et al. 1999, Fuhrmanet al. 1996, Cheng and Chang 1996, Qiu 1997, Jagannathan and Talluri2002) and without neural and fuzzy logic (Bonomi and Fendick 1995,Benmohamed and Meerkov 1993, Jain 1996, Mascola 1997, Izmailov 1995,Bae and Suda 1991, Liu and Douligeris 1997, Bonomi et al. 1995, Jagannathanand Talluri 2000). In terms of feedback control, the congestion control isviewed as changing the source rates and regulating the traffic submittedby these sources onto the network connections while ensuring fairness.Most of these schemes are based upon a linear model of the ABR buffer.Each ABR buffer has a corresponding congestion controller, which sendscongestion notification cells (CNC) back to the sources. These feedbackcontrol schemes from the literature were based on buffer length, bufferchange rate, cell loss rate, buffer thresholds, and so on.The most simple form of congestion control scheme is the binary feedbackwith FIFO queuing. Here, when the buffer occupancy exceeds apredefined threshold value, the switch begins to issue binary notificationsto the sources and it continues to do so until the buffer occupancy fallsbelow the threshold. This is referred to as thresholding in this chapter.Unfortunately, a FIFO queue may unfairly penalize connections that passthrough a number of switches and therefore more advanced schemes,including explicit rate feedback schemes such as ERICA have appearedin the literature. Here, the feedback would provide an explicit rate to thesources via RM cells and these schemes ensure a certain degree of fairness.However, these schemes are not supported by rigorous mathematicalanalyses to demonstrate their performance, in terms of QoS, and rely onlyon some simulation studies. The author, in Benmohamed and Meerkov(1993), has reported that certain congestion control schemes even resultedin oscillatory behavior. In Benmohamed and Meerkov (1993), a stablefeedback congestion controller is developed for packet switching networksbased on the linear control theory notion. Although a stable operationis obtained with a linear controller, the network traffic rate is notestimated and the impact of the self-similar nature of network traffic ontheir controller design is not considered in the analysis.In this chapter (Jagannathan and Tohmaz 2001, Jagannathan and Talluri2002), the network traffic accumulation at the switch is considered asnonlinear. A novel adaptive multilayer NN congestion controller schemeis developed using a learning methodology, where an explicit rate for each