2005 IEEE 16th International Symposium on Personal, Indoor and Mobile Radio Communications Clustered-loss Retransmission Protocol over Wireless TCP1 Fanglei Sun and Victor O.K. Li Dept. of Electrical and Electronic Engineering The University of Hong Kong, Pokfulam, Hong Kong, China Abstract Transmission Control Protocol (TCP) performs well in traditional wired networks where the packet loss rate is low. However, in heterogeneous wired/wireless networks, the high packet loss rate over wireless links may result in excessive invocation of the congestion control algorithm, thus deteriorating the performance of TCP. In this paper, a novel localized link layer retransmission protocol, called Clustered-loss Retransmission Protocol (CLRP), is proposed. CLRP consists of three protocol components, namely, TCP-FH deployed on a fixed host, TCP-MH deployed on a mobile host and CLRP-BS deployed on a base station. CLRP can provide not only explicit distinction between congestion and packet corruption losses, and effective multiple wireless loss information for retransmissions, but also better retransmission control for wireless losses. Thus it is well suited to wireless networks, in which packet loss and bursty packet corruption is a serious problem. Moreover, CLRP does not require any modifications to TCP deployed on fixed hosts. I. INTRODUCTION During the past few years, the proliferation of competing technologies and service network models has accelerated the growth of the wireless Intemet. The congestion control algorithms embedded in TCP work well in wired networks in preventing congestion collapse. However, in heterogeneous wired/wireless networks, TCP regards both wired and wireless packet losses as an indicator of network congestion, and thus, TCP and its variations, such as TCP Reno, TCP Newreno and TCP SACK, will invoke the congestion control algorithm although the losses may not be caused by congestion. Therefore, how to allow TCP to distinguish between the losses due to congestion and due to packet corruption in a timely fashion has become the crux of the research on wireless TCP. Several approaches to address this problem have been proposed to improve TCP performance over wireless networks [1]. These approaches include end-to-end mechanism like Veno [2], split connections mechanism like M-TCP and localized link layer mechanism like Snoop. This paper focused on the localized link layer solution and the proposed Clustered-loss Retransmission Protocol (CLRP) can perform better than the existing localized link-layer approaches. We consider the transmission between a fixed host (FH) and a mobile host (MH) relayed through a base station (BS). The Snoop protocol (Snoop) [3,4] installed at the link layer of a BS monitors the packets and ACKs in both MH to FH and FH to MH directions. For a transmission from FH to MH, Snoop stores the packets arriving at the BS and arranges local retransmissions based on the type of ACKs and local timers. 1 This research is supported in part by the Areas of Excellence Scheme established by the University Grants Committee of the Hong Kong Special Administrative Region, China, under project No. AoE/E-O1/99. Soung C. Liew Dept. of Information Engineering The Chinese University of Hong Kong, Hong Kong, China From MH to FH, Snoop adds explicit loss notification (ELN) [5], namely, setting the value of one bit in the six reserved bits included in a TCP header, thus allowing MH to distinguish congestion losses from wireless random losses. However, Snoop can only provide single packet loss information within one RTT (round-trip-time). Under high loss rate wireless environment, Snoop does not work well because it mimics the TCP error recovery mechanism, which is not very robust under harsh error conditions. In bursty traffic network, the lack of explicit and accurate information in Snoop degrades the bandwidth utilization sharply. Furthermore, Snoop offers great improvement in wired-cum-wireless networks, i.e. the transmission is from a fixed host to a mobile host. But when used in wireless-cum-wired or wireless-cum-wireless networks, Snoop is regarded as ineffective [6]. Clustered losses result from bursty multiple packet losses. When multiple packets are lost in a TCP window and within one RTT, the congestion window size will be reduced continuously, degrading the throughput nearly to zero. As a result, timeout is used by TCP to recover packet losses. To overcome this defect, a selective acknowledgment (SACK) mechanism is proposed in RFC 2801 [7]. In TCP SACK, several SACK blocks are used to inform the sender about all the segments that have been received successfully, which allows the sender to retransmit only the lost segments. Each SACK block consists of the beginning and the ending sequence number of a consecutive packet block received by the sender, and thus the holes between the SACK blocks are regarded as lost packets. However, TCP SACK will also cause the following problems: 1) SACK blocks piggy-backed in ACKs take up much space left in the TCP option; 2) SACK blocks transmitted between FH and MH decreases the transmission efficiency, particularly for the transmission with small TCP packet size. Furthermore, the mutual interference between TCP SACK and Snoop when processing bursty losses on wireless links is also shown in [8]. Therefore, it is impractical to solve the problem of clustered losses over wireless networks by using a combination of TCP SACK and Snoop. If TCP SACK is used in the BS directly, it may violate the end-to-end semantics of TCP. Recently, much research has been focused on designing a new ACK [9] for wireless TCP. Unfortunately, it encounters the same problems as TCP SACK. In this paper, for typical heterogeneous wired/wireless networks consisting of three components, namely, fixed host 978-3-8007-2909-8/05/$20.00 ©2005 IEEE 1 505