Allowing Short-Lived TCP Sessions to Ramp-UP in Broadband Wireless Networks Toktam Mahmoodi, Vasilis Friderikos, Hamid Aghvami Centre for Telecommunications Research, King’s College London, Strand, London WC2R 2LS, UK {toktam.mahmoodi, vasilis.friderikos, hamid.aghvami}@kcl.ac.uk Abstract—In this paper, we propose a dynamic OFDMA based subcarrier/power allocation which aims to protect the newly established TCP connections. The proposed allocation utilizes an estimation of the current congestion window of TCP connections to allocate resources so that vulnerable flows with small congestion windows avoid entering packet recovery via retransmission time out. In addition, for the long-lived TCP flows, the theoretical TCP throughput which can be accomplished in their end-to-end path is used as their desired achievable data rate on wireless link. Such a technique can be of significant importance since due to its popularity, TCP is commonly used over the broadband wireless networks. I. I NTRODUCTION As the Internet grows both in terms of the number of users and diversity of applications, providing fair and efficient allocation of the available network resources becomes increas- ingly challenging. The Transmission Control Protocol (TCP) is the default transport layer protocol used in the Internet to provide reliable end-to-end communications. TCP carries 95% of today’s Internet traffic and 80% of the total number of the flows in the Internet among which a large majority are short-lived flows [1][2]. However, TCP exhibits a number of shortcomings when the underlying wireless medium deviates from the reliability of the wired medium for which TCP was originally designed to serve [3]. A new TCP connection initializes in slow start phase, where the congestion window (cwnd) is small and packet loss can be a dramatic event. The occurrence of a a packet loss in the newly established TCP connection, can lead to the occurrence of the Retransmission Time Out (RTO), causing delay to the data transmission in the order of seconds. The repetition of such an event exponentially increases the RTO due to the TCP back-off, resulting in a large aggregate delay at the very outset of the session, which may cause the user to prematurely terminate the session. On the other hand, in the slow start phase cwnd is small, thus a TCP flow is not sensitive to the allocated data rate. Unlike long-lived TCP connections, the performance of short-lived connections is highly dependent on the slow start phase. Therefore, various approaches are discussed in the literature to either shorten the slow start phase or improve the performance of TCP in this phase. For example, “Quick- Start” for TCP [4] allows TCP to capture its desired sending rate by quickly increasing its cwnd. Another research work presented in [5] utilizes the cross-layer information on the available bandwidth to increase its cwnd accordingly in the slow start phase. In all these works, a bottom-up approach is used. Alternatively, we advocate the use of a top-down approach as well, i.e., the resource allocation scheme at the wireless link-layer protects the TCP flows in the slow start phase. This not only allows for the performance of short- lived TCP connections to be increased, but also provides an opportunity for the newly established TCP connections to grow their cwnd and to utilize the available wireless bandwidth. After crossing a certain threshold, a TCP flow will enter the congestion avoidance phase, where it requires sufficient resources in order to maintain its cwnd and further increase its data rate. In this phase, where the cwnd is larger, the main constraint is the allocated wireless channel to the correspond- ing TCP flow, even with the expense of higher error rates, since low channel rate inevitably results in packet loss due to congestion. The TCP-aware channel allocation algorithm in CDMA networks proposed in [6] adapts the channel rate in response to the TCP sending rate. Another approach presented in [7], providing extra protection for TCP flows that have a cwnd smaller than the network bandwidth-delay product by using more robust modulation or higher transmission power for them. To this end, with respect to the evolution of TCP cwnd, we define a state diagram for each TCP flow such that being in any state requires a different level of protection, and data rate. Thus, a TCP-aware resource allocation algorithm for OFDMA-based access networks is proposed that adapts its target Bit Error Rate (BER) and minimum required data rate per flow according to the state of the corresponding flow. This resource allocation scheme attempts firstly, to protect the newly established TCP connections, and secondly to provide a more balanced performance towards TCP throughput that is not only an optimal resource allocation for each TCP flow, but also is a fairer distribution among TCP flows. The data rate constraint is based on the estimated current cwnd of TCP for short-lived TCP connections, and on the closed form expression of TCP steady state throughput [8] for the long- lived TCP connections. In addition, we study the performance of the proposed method, for various combination of short-lived and long-lived TCP flows and extensive simulation results are presented. Although all competing end-to-end flows in the investigations of this work are TCP flows, other end-to-end flows such as UDP can also be assumed among the active 978-1-4244-5626-0/09/$26.00 ©2009 IEEE