TCP performance enhancement over wireless networks (2007)
Type of ContentTheses / Dissertations
Thesis DisciplineElectrical Engineering
Degree NameDoctor of Philosophy
PublisherUniversity of Canterbury. Electrical and Computer Engineering
AuthorsJayananthan, Aiyathuraishow all
Transmission Control Protocol (TCP) is the dominant transport protocol in the Internet and supports many of the most popular Internet applications, such as the World Wide Web (WWW), file transfer and e-mail. TCP congestion control algorithms dynamically learn the network bandwidth and delay characteristics of a network and adapt its performance to changes in traffic so as to avoid network collapse. TCP is designed to perform well in traditional wireline networks with the assumptions that packet losses are mainly due to network congestion and random bit error rate (BER) is negligible. However, networks with wireless links suffer from significant packet losses due to random bit errors and handoffs. Hence TCP performs poorly in networks with wireless links because it treats any packet loss in the network to be a result of network congestion and slows down its transmission rate, or even cause the TCP sender to experience unnecessary timeouts, further reducing its performance. The development of advance wireless networks, such as WiFi, UMTS and WiMAX, make it necessary to find ways to improve TCP's efficiency and resource utilization, as well as improve the user's experience and reduce latency times. In order to find effective solutions to this effect, packet losses across wireless links should be distinguished from congestion related packet losses. In this thesis, we concentrate on two main strategies for enabling the TCP congestion control mechanism to determine the cause for a packet loss. One is a proxy-based mechanism that monitors the radio network interface and sends radio network feedback (RNF) to the TCP sender with the status of the wireless link. The other one is an end-to-end mechanism, in which the packet error pattern is used as the system metric to fine-tune the congestion control mechanism. It also presents an analytical model of TCP with enhanced recovery mechanism for wireless environments. In a proxy-based mechanism, TCP sender is explicitly informed of any effects caused by wireless links. However, the implementation technique is network dependent. We have proposed and developed three proxy-based schemes; the radio network feedback (RNF) scheme over an 802.11 WLAN network, the radio network controller (RNC) feedback over a UMTS network and a wireless enhancement proxy (WENP) over both the 802.11 WLAN and UMTS networks. The RNF scheme is introduced at the 802.11 WLAN base station that monitors the TCP packet flows over the wireless links, detects wireless packet losses and provides feedback to the TCP sender using one of the TCP header reserved control bits, called RNF flag. TCP Reno is modified to utilize the radio network feedback to distinguish the losses due to wireless effects form the congestion and fine-tuned to perform wireless enhanced fast retransmit and fast recovery mechanisms. The RNF scheme is implemented using the OPNET tool, and the simulation results show that the TCP performance is significantly improved. The RNC feedback mechanism, similar to the RNF scheme, is developed and implemented in a UMTS network. The GPRS Tunneling Protocol (GTP) layer of the UMTS Radio Network Control (RNC) protocol stack was modified to detect and notify the TCP sender of the wireless packet losses, which is the main difference between the RNF and RNC mechanisms. The simulation results shows that the RNC feedback mechanism significantly improves the TCP performance compared to that of standard TCP over UMTS. The wireless enhancement proxy (WENP) is developed to minimize spurious TCP timeouts over wireless networks and implemented in both 802.11 WLAN and UMTS networks. WENP extends the proposed RNF and RNC feedback mechanisms to detect both wireless packet losses and large delays across the wireless link, and to notify the TCP sender of these events with the aid of two reserved bits in the TCP header. TCP Reno is further modified to utilize the WENP feedback to distinguish both wireless packet losses from congestion losses and spurious timeouts from normal timeouts. It is also fine-tuned to perform both the wireless enhanced fast retransmit and fast recovery mechanism and the timeout mechanism. The simulation results demonstrate that the proposed scheme markedly improves the TCP performance compared to that of standard WLAN and UMTS implementations. An end-to-end early packet loss recovery (EPLR) mechanism that modifies the TCP Reno fast retransmit algorithm to detect packet losses early and to speed up the packet recovery process to reduce the number of TCP timeouts over networks with heavy packet losses, such as wireless networks is also presented. TCP Reno with EPLR scheme is implemented in a UMTS network and its performance is compared with that of TCP Reno and New Reno. Simulation results shows that Reno with EPLR improves the TCP performance and application response time significantly compared to that of both Reno and New Reno by reducing the TCP timeouts, which is the main cause of degradation of the TCP performance in a wireless environment. Finally, we develop an analytical TCP throughput model with enhanced TCP Reno fast retransmit algorithm to avoid timeouts. The model captures the TCP fast retransmit mechanism and expresses the steady state congestion window and throughput as a function of network utilization factor, round trip time (RTT) and loss rate. Another new feature added to the model is dynamic adjustment of the congestion window size depending on the packet drop rates. This speeds up the packet recovery process and reduces the number of TCP timeouts over networks with heavy packet losses. The proposed model is implemented over a UMTS network and its performance is compared with that of TCP Reno. Simulation results show that the proposed model reduces the TCP timeouts and improves the TCP performance compared to that of TCP Reno. It is also found that the model provides a very good match to the steady-state congestion window behavior.