Wireless Ad Hoc and Sensor Networks

Congestion Control in ATM Networks and the Internet 109(Jagannathan and Talluri 2002, Benmohamed and Meerkov 1993, Chenand Chang 1996, Widmer et al. 2001, Floyd 2001, Jain 1996, Kelly et al.2000, Floyd and Fall 1999, Low et al. 2002, Jagannathan and Talluri 2002,Chandrayana et al. 2002) have been proposed on congestion control. Some(Jagannathan 2002, Benmohamed and Meerkov 1993, Cheng and Chang1996) are based on nonlinear system theory (Low et al. 2002, Jagannathanand Talluri 2002, Chandrayana et al. 2002), whereas others are based onheuristics and past experience (Floyd 2001). Most schemes adjust the inputrates to match the available link capacity (or rate) (Jain 1996, Lakshmanet al. 1999) or carefully admit new traffic (Kelly et al. 2000, Mortier et al.2000). A comparison of the performance of the TCP-based end-to-endcongestion control variants are presented in (Fall and Floyd 1996).The basis of the congestion control scheme in the Internet (Floyd 2001)lies in additive increase and multiplicative decrease (AIMD) algorithm,halving the congestion window for every window containing packet loss,and otherwise increasing the congestion window roughly one segmentper round trip time (RTT) (Floyd 2001, Jain 1996). The second componentis the retransmit timer, which includes the exponential backoff when atransmitted packet itself is dropped. Another component is the slow-startmechanism for initial probing for available bandwidth instead of sendingit at a high rate. The fourth component is acknowledgement (ACK) clocking,where the arrival of acknowledgments at the sender is used to clockout the transmission of new data. Within this general congestion controlframework of slow start, AIMD, retransmit timers, and ACK clocking,there is a wide range of possible behaviors.Unfortunately, the AIMD-based congestion control scheme appears tohave several problems:1. It does not guarantee fair share to all the users (elastic and inelastic)(Widmer et al. 2001).2. It is not proven analytically to converge and the stability of theentire system is not shown.3. The loss of packets in the Internet may be construed as congestion.4. The initial congestion window is currently selected heuristically.5. It is a reactive scheme; a predictive scheme is preferable due tofeedback delays. Finally, the congestion control problem in theInternet is exacerbated as the Internet is increasingly transformedinto a multiservice high-speed network (Widmer et al. 2001).Random early detection (RED) scheme for controlling average queuesize is proposed so that unnecessary packet drops can be avoided (Floyd2001). Further, explicit congestion notification (ECN) bit is introduced,which, when built upon queue management, allows routers the option of