PUBLISHED IN: PROCEEDINGS OF THE IEEE AINA 2006 1 Adaptive Timeout Policies for Wireless Links George Xylomenos and Christos Tsilopoulos xgeorge@aueb.gr and p3000173@dias.aueb.gr Mobile Multimedia Laboratory Department of Informatics Athens University of Economics and Business Patision 76, Athens 104 34, Greece Abstract— A considerable body of evidence indicates that the use of reliable link layer protocols over error prone wireless links dramatically improves the performance of Internet protocols and applications. While traditional link layer protocols set their timeout values assuming that they fully control the underlying link, some wireless networks allow multiple link layer sessions to co-exist over the same link. Since the optimal timeout values for a reliable link layer protocol depend on the available bandwidth, with dynamic link sharing such a protocol should ideally adapt its timeout values accordingly. We have thus designed an Adaptive Selective Repeat protocol that modifies its timeout values based on the policy used by TCP. We compare the performance of Web Browsing over Selective Repeat when using our adaptive timeout scheme with a range of parameters, against a manually tuned fixed timeout version. Our measurements show that these adaptive timeout policies outperform the fixed one, regardless of the level of contention, and that the best adaptive timeout policy in this setting is not the one used by TCP. I. I NTRODUCTION While wireless networks are becoming increasingly popular as access network to the Internet, many researchers have found that the error prone nature of wireless links dramatically degrades the performance of popular applications such as Web Browsing [1]. A direct way to hide wireless link deficiencies is to employ a reliable link layer protocol; such protocols have been shown to improve the performance of higher layer protocols and applications, even when unaware of the nature of the higher layers used [2]. Traditional link layer protocols assume that they have full control of the underlying link when choosing their parameters, including timeouts. Due to the higher bandwidth offered by emerging wireless networks however, it is becoming common for multiple users and/or applications to dynamically share a single wireless link. For example, in the Universal Mobile Telecommunications System (UMTS) a single physical channel is shared among the link layer sessions of different users and/or applications. Even in links controlled by a single user, it is desirable for multiple such protocols to co-exist, so as to serve different applications. In this paper we examine the impact of link layer timeout policy on application performance over shared links. In Sec- tion II we provide background information on Internet protocol and application performance over wireless links. In Section III we discuss the problems faced by a Selective Repeat protocol in this environment and propose an adaptive timeout scheme. Section IV describes our simulation setup for the performance evaluation. In Section V we show Web Browsing performance when operating in isolation over a wireless link, while in Section VI we show Web Browsing performance when con- tending for the link with a Media Distribution application. Our measurements show that the adaptive timeout policies outperform the fixed one. II. BACKGROUND AND RELATED WORK The Internet Protocol (IP) offers an unreliable packet delivery service. Many real-time applications use the User Datagram Protocol (UDP) for direct access to IP, handling error, flow and congestion control themselves. Most other applications prefer delegating these tasks to the Transport Control Protocol (TCP) which offers a reliable byte stream service. TCP segments the application data stream into packets and reassembles it at the receiver. The receiver generates acknowledgments (ACKs) for segments received in sequence, returning duplicate ACKs for out of sequence ones. The next unacknowledged segment is retransmitted after receiving 3 duplicate ACKs or when a timeout occurs. Since wired links are extremely reliable, TCP assumes that all packet losses are due to congestion, thus after a loss it abruptly reduces its transmission rate, and then gradually increases it so as to probe the network. Unfortunately, this means that losses due to wireless errors are mistaken for congestion signals, causing TCP to dramatically reduce its transmission rate [1]. Many modifications have been proposed to improve TCP performance over wireless links, but they all have two drawbacks: they require modifications to end hosts throughout the Internet and they can only retransmit lost data over the (possibly long) end-to-end path. A simpler approach is to employ a reliable link layer protocol over the wireless link to locally hide wireless errors from TCP. An early proposal customized to TCP snoops inside each TCP stream at the access point bridging the wired and wireless parts of the path and retransmits lost segments when duplicate ACKs arrive, hiding them from the sender to avoid end-to-end recovery [3]. However, later work shows that TCP performance can be enhanced even with standard, TCP unaware, reliable link layer protocols [2]. It is important to note that not all Internet applications require full reliability. While TCP based applications perform better with reliable link layers, delay sensitive UDP based applications often prefer faster, albeit limited, error recovery.