Optimization of TCP Over Wireless Networks

Transcription

1 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 11 No: Optimization of TCP Over Wireless Networks Mohammad Usman Ali Khan Peshawar College of Engineering, Pakistan Shahid Khan Politecnico di Torino, Italy Yasir Ali Rasheed National Vocational & Technical Training Commission (NAVTTC), Pakistan Abstract As mobile devices are popular portals for internet users; they have some basic design limitations of the TCP having congestion control mechanism. But in mobile wireless networks, non congestion related packet losses due to varying signal power inherent with mobility and handover between base-stations are dominant. Due to the misinterpretation of such non congestion related packet losses optimum performance is compromised. This paper provides an overview of the challenges and suggests optimization techniques for TCP in wireless network environments. The consequence is a wireless enhancement scheme for TCP called Congestion Coherence. By several investigations and statistical analysis, we identified that the performance degradation of wireless TCP is mainly due to the timeout resulting from frequent packet losses. This scheme determines the cause of packet loss by using the sequential coherence of packet marking, it uses three-level ECN signals to detect congestion loss and to reduce number of timeouts established on the perception that congestion neither happens nor disappears abruptly. Simulations show that the performance and efficiency of this technique is better than the existing techniques. Index Term-- ECN, TCP, ECN-Echo I. INTRODUCTION There has been a mammoth growth of wireless networks in the past decade, which has been driven mainly due to the need of accessing the Internet any time, any where [7], and also due to the massive use of mobile telephony for communications. Wireless communications is available in many forms and spans a great range of throughput and distances. The world of wireless includes wireless LANs, fixed microwave links, digital dispatch networks, one-way and two-way paging networks, satellite links, diffused infrared, laser-based communications, the Global Positioning System, keyless car entry and a lot more. Wireless communication equipment is easily available for use with computers in the past decade. Wireless LAN is commonly implemented with IEEE standard [9]. Wider area coverage is made possible by utilizing the existing mobile phone infrastructure for data transmission in the form of, for example, GPRS [3, 4]. The integration between packet data and voice communications is further evolved with the deployment of 3G/UMTS [10]. This system is specifically designed to carry packet data, video and voice communications with much higher capacity than previous wide area coverage networks. Parallel to this activity is the expansion of the Internet, which has grown tremendously since the introduction of the World Wide Web. In just a few decades, the number of networks has increased from tens to hundreds of thousands [1]. Although wireless networks provide mobility and convenience to users but still conversations may be cut off when mobile subscribers travel between cells. Wireless data connections have high bit error rates, low bandwidth and long delays. The performance of the Internet protocols is much lower in wireless networks [6, 8] because TCP works less efficiently in wireless networks [11] as compared to [2, 5] traditional wired networks with stationary hosts, it assumes congestion in the network to be the primary cause for packet loss and unusual delay. In a wireless network most losses are not caused by congestion but by low quality of the wireless link. Terminal mobility, which is supported by many wireless networks, may also result in data loss [11]. The bandwidth of the wireless media is limited and shared by many users so its throughput is lower than that of wired media. The competition for the shared medium causes access contention. If not coordinated properly, the access contention can result in low efficiency and unpredictable channel access delay. The, wireless media is prone to transmission errors so Forward Error Correction (FEC) and Automatic Retransmission Request (ARQ) are used to recover transmission errors. In wireless WAN systems, the average bit error rate after FEC can be as high as Link layer ARQ can reduce the bit error rate perceived by the high protocol layers to a magnitude of 10-3 [12]. The remaining transmission errors are usually perceived as packet losses by upper protocol layers. The wireless media is asymmetric. In cellular phone system, the radio connection consists of a downlink from the base station to the mobile hosts, and an uplink from the mobile hosts to the base station. The two links are not

2 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 11 No: symmetric. Most existing networking protocols were designed for wired links and they improved the efficiency in wired networks but not apply to wireless links. With the rapid development of wireless networks, identifying the causes of poor performance and enhancing the current protocols has become an urgent task. II. TRANSMISSION CONTROL PROTOCOL OVERVIEW TCP [2, 13, 14] is most widely used connection-oriented protocol designed for an unreliable network layer. It provides reliable data delivery and strong congestion control mechanism above the network layer and is foundation of today s Internet. TCP algorithm is tasked with the end-to-end integrity of the transfer of user data. To establish a TCP connection, the source transmits a SYN request message segment, the destination replies with its own SYN-ACK segment; then the source answers with an ACK segment. Data can be transmitted in both directions once the connection is established. This establishment of connection is called a three-way handshake. TCP is a reliable transport protocol and uses an acknowledgement mechanism to confirm the safe delivery of data. To ensure connectivity, each byte to be transmitted is assigned a unique sequence number and the receiver sends ACK upon receipt of a segment. A TCP acknowledgment can be a separate packet, or can be piggy backed onto a data packet in the reverse direction. TCP waits for nearly 500 ms before acknowledging a received segment [13]. A. TCP Congestion Control Mechanisms Congestion occurs when the data exceeds the network s capacity and the queue at the buffer of the router increases and data packets suffer an extended delay. If congestion continues, the queue size exceeds the buffer size and cause buffer overflow and packets start dropping, and have to be retransmitted later from the source. Normally some packets that have already passed the congested router are also retransmitted and increases delay which also consumes bandwidth. Figure 2.1 [15] illustrates the delay and throughput of a network in response to increasing load. Fig. 2.1 [15]. Load versus delay and throughput TCP controls congestion rather then avoid it in the first place. TCP gradually increases the traffic on the network to find the point at which congestion occurs to find the available bandwidth and then backs off from this point. This mechanism is known as congestion control. An alternative, which is called congestion avoidance, is to predict when congestion is about to happen and then reduce the transmission rate just before packets start being discarded. B. Congestion Feedback Mechanisms The TCP algorithm deployed at the source and destination is only part of the congestion control effort. The basis of any congestion control solution is the congestion feedback from the network. The most primitive of them is the packet dropping, in which the network simply drops incoming packets. The most common packet dropping policy is the Tail Drop. The second approach used by several protocols is to send congestion control messages incase of congestion. Another approach, called explicit notification, is to mark congestion information in the header of passing packets. When these packets reach the receiver, the attached info is sent back to the sender through acknowledgment packets. C. Wireless TCP Enhancements Wireless TCP enhancement has two categories. The first category attempts to hide wireless losses from TCP so existing TCP algorithms do not need to change. This category can be further divided into split connection approach, link layer protocols, and transport layer protocols. The second category conveys wireless losses explicitly to TCP and lets TCP decide what congestion action to take. III. PROBLEMS AND REQUIREMENTS OF WIRELESS TCP ENHANCEMENTS The performance degradation of wireless TCP is caused by TCP s inability to distinguish between wireless and congestion losses. When TCP is utilized on wireless networks, the effective transmission rate is reduced to half and the throughput of the link is reduced severely which causes heavy TCP performance degradation. Wireless TCP is heavily suffered from timeout, which is the result of multiple losses in one window and the chance of having multiple packet losses is significantly increased, causing severe performance degradation. Timeout is the leading and dominant contributor of performance degradation. In addition, we realize that current wireless TCP enhancements fail to improve TCP s timeout behavior and therefore exhibit lower performance in many cases. Several enhancement schemes have been presented but many of these are not practical in terms of assumptions, implementation and applicability. As criteria to evaluate these enhancements, we establish a number of necessary requirements for wireless TCP enhancements. A. Implementation Requirements Here is summary of requirements that are desirable in practice, listed in the order of importance. 1) TCP Semantics Requirement First, the enhancement scheme must maintain the end-to-end TCP semantics. All correctly received packets must be acknowledged. 2) Local Modifications Requirement All modifications to the existing TCP algorithm must be local. A wireless link added to the network should not change the entire network.

3 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 11 No: ) Encrypted Traffic Requirement Security is becoming a prime need; a large portion of traffic is encrypted. Some encryption protocols encrypt IP payload so that intermediate nodes cannot see what is being carried. 4) Two-way Traffic Requirement Since the traffic is affected by wireless errors in both directions, so in order to fully utilize the wireless bandwidth; the enhancement scheme must apply to two-way traffic. 5) Intermediate Wireless Link Requirement The enhancement scheme should also be applicable to intermediate wireless links In addition to the above stated requirements, it is required to control the traffic, buffering and processing overhead. Table 3.1 summarizes the main characteristics of these enhancements. It includes the changes made in these enhancement schemes, their complexity and performance. TABLE 3.1 COMPARISON OF WIRELESS TCP ENHANCEMENT SCHEMES I-TCP Control Multiple Connection Acks ELN DDA Snoop Modify BS Yes No Yes Yes No Yes Modify MH No Yes No No Yes No Modify FH No Yes Yes No No No BS Buffering Heavy No Heavy No Light Heavy Two way traffic Yes No No No No No Interpret losses No Probably Yes Yes No Yes Interpret Dup Acks No Probably Yes Yes Yes Yes Delay variations Small Large Small Large Small Small Complexity Simple Moderate Moderate High Simple High End-to-end semantics No Yes Yes Yes Yes Yes IV. SIMULATION AND VALIDATION OF CONGESTION COHERENCE FOR WIRELESS TCP ENHANCEMENT Current proposals use packet losses to deliver congestion feedback and fail to improve the TCP s timeout behavior over wireless links. A good enhancement scheme must stop using packet losses as the congestion feedback mechanism. A transmission rate drop scenario is illustrated in Figure 4.1. Before any transmission error happens, the link can support a TCP transmission window size of six packets. When the sender starts transmitting packets at this rate, a packet gets lost on the wireless link due to a wireless error. When the destination receives an out-of-order packet, it follows TCP congestion control algorithm to send back duplicate ACKs. Fig Window reduction due to wireless losses Numerous enhancement schemes have been proposed to improve performance of TCP over wireless links. These schemes suggest modification at the fixed source, the mobile host or the base station. Some enhancements try to hide the lossy characteristic of wireless link from TCP by performing buffering and retransmission at the base station while others use extra traffic to determine the network congestion status. We have reviewed and evaluated these methods, and concluded that these enhancements make unrealistic assumptions, often fail to meet certain implementation requirements, or apply only to specific network configurations. TCP traces reveals that wireless errors can degrade TCP performance in three ways: End-to-End Retransmissions Whenever a packet is lost, it is retransmitted by TCP. If an end-to-end retransmission is done, the extra packets added to the traffic and reduce the number of packets that can be sent through the bottleneck link. Unnecessary Slowdowns TCP algorithm treats packet loss as an indication of congestion, whenever a wireless loss is retransmitted end-to-end, and thus reduces the congestion window size. Severe performance degradation is caused due to unnecessary slowdowns because the effective transmission rate is cut to half. Timeouts Existing TCP algorithms can recover a single packet loss in a window using duplicate acknowledgments. Two or more packet losses due to congestion or wireless error in the same window will result in a timeout and takes many round trip times (RTT) to bring the transmission rate to a normal value. A. Proposed Approach In order to improve TCP performance, the proposed enhancement method should eliminate the above three degradations. The first degradation can easily be taken care of by local link layer retransmissions. The second degradation can be removed by distinguishing between wireless and congestion losses. For the final degradation, we need to stop using packet loss as a mechanism to fix the TCP window. All existing solutions eliminate only the first two degradations and use packet losses to deliver congestion feedback and therefore, the third degradation is inherent and unavoidable. We deploy Explicit Congestion Notification (ECN) for congestion control mechanism and propose a solution to distinguish wireless and congestion losses. ECN avoids congestion losses by making use of early congestion warnings. By avoiding congestion losses we achieve two performance benefits. In this way it avoids the end-to-end retransmissions, as well as timeouts caused by multiple losses in a window. This is how we can eliminate all the three performance degradations of TCP over wireless links. Our proposal consists of the following three parts: Assume that all network routers are three-level ECN capable. Implement local link layer retransmission on the wireless link to avoid end-to-end retransmission of the wireless losses.

4 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 11 No: Use congestion coherence with three-level ECN to determine the cause of packet loss. 1) Explicit Congestion Notification The idea of delivering feedback on congestion status by marking packet header at the router can be traced to the DEC bit scheme [17]. Explicit congestion feedback delivers a direct and faster signal to the sender compared to timeout and duplicate ACKs. ECN has proven to be quite effective in avoiding unnecessary packet drops. ECN uses four bits for ECN capability negotiation and feedback delivery; two in the IP header and two in the TCP header. When the queue size at the router exceeds the threshold and the incoming packet is labeled ECN-capable, the router marks the packet as congestion experienced. When this packet arrives at the destination, the receiver copies the congestion experience bit to the ECN-Echo bit in the next acknowledgment packet and thus delivers a congestion notification back to the sender. The sender reduces its congestion window, upon receiving the ECN-Echo, to mitigate the congestion. 2) Local Link Layer Retransmission A lost packet has to be retransmitted, either end-to-end or over the local wireless link. Local link layer retransmission not only avoids extra traffic in the network but also reduces the delay in comparison to the end-to-end retransmission. In such case the source may timeout and triggers an endto-end retransmission before the acknowledgment from the local retransmission is received. When designing any link layer retransmission scheme, it is crucial to transmit the failed packets first to reduce the retransmission delay. One way of implementing this is by deploying the insert from front strategy. Whenever a packet is noted to be lost; the failed packet is inserted into the transmission queue s front by the link layer, and transmitted as soon as the media gets available. When a lost packet is retransmitted, the destination will receive out-of-order packets and may respond with duplicate ACKs. Both wireless losses and congestion losses cause out-of-order packet delivery, and create spaces in the packet sequence number but with different consequences. When a retransmission arrives, any space induced by a wireless loss will be covered, but incase of a congestion loss if an end-toend retransmission is not triggered or a timeout does not occur, the created space will not be covered. To handle this issue; the receiver must wait for the retransmission when it detects that the space is caused by a wireless loss and not by congestion loss. 3) Congestion Coherence Table 4.1 illustrates two loss cases. Packets 12 and 37 are lost. E means the packet is marked as Congestion Experienced and blank means the packet did not experience any congestion. TABLE 4.1 ARE THE PACKETS LOST DUE TO CONGESTION OR DUE TO TRANSMISSION ERROR? Seq. No. Mark Seq. No. Mark E E E 12 Lost 37 Lost E E Packet 37 is a congestion loss and packet 12 is a transmission loss. This observation is based on congestion coherence of ECN markings. Congestion does not totally fade out after a packet is dropped. As a result, marked packets are commonly received before and after packets lost due to congestion. Figure 4.2 [18]. Fig. 4.2 [18]. Congestion coherence In contrast, transmission errors normally happens independent of congestion. Congestion coherence can be used to distinguish congestion losses from wireless losses. We define the coherence context of packet x as packets [x 1, x + 1, x + 2], the coherence context is to be considered only if any of the context packets are marked. If the context is not marked, it is most likely a wireless loss. Duplicate ACKs should be held to avoid invoking fast retransmit and congestion control at the source. The above stated concept can also be implemented to wireless case. Every time a duplicate acknowledgment is received by a wireless source, it inspects the coherence context for an ECN-Echo. If an ECN-Echo is seen, the wireless sender assumes that a congestion loss caused the duplicate acknowledgments, it invokes the congestion control. Incase no ECN-Echo is found, the wireless sender assumes that a wireless loss has caused the duplicate acknowledgments, it simply ignores the duplicate acknowledgments until the local retransmission succeed. All the modifications in the existing TCP algorithm for this approach (Congestion Coherence) are based on the procedure discussed above are made in the wireless end. Thus by implementing Congestion Coherence at the wireless end, lossy characteristic of the wireless link can be hidden from the other end. No modifications are needed at the wired end (including the base station) or the intermediate routers. Details of congestion coherence algorithm are described by Raj Jain and Chunlei Liu in [18]. 4) Proposed Scheme: Three-Level ECN Congestion Coherence scheme, we discussed above uses ECN to distinguish wireless and congestion losses. Current ECN uses two bits in the IP header [16, 17]. The first bit is called ECT (ECN Capable Transport) bit. It is set if both sender and receiver are ECN capable and wish to use ECN. The second bit is called the CE (Congestion Experienced)

5 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 11 No: bit. If the ECT bit is set in a packet, the router can set the CE bit to indicate congestion. This scheme is not efficient, because it uses two bits indicate three states: not ECT; ECT and congested; and ECT and not congested. Since two bits can represent four states, a more efficient scheme is possible. Based on the above observation and the proposal in previous subsection, a new scheme has been presented [19], which uses three levels of congestion with ECN signals to detect congestion losses. We propose to change the ECN field in the IP header and ECN-Echo field in the TCP header. The ECN field consists of two bits and the ECN-Echo field consists of two bits. The ECN field is set by the source and changed by the intermediate routers; the ECN-Echo field is set by the receiver. The proposed bit patterns are summarized in Table 4.2. With these bit patterns, network routers can indicate three levels of congestion with ECN signals: no congestion, mild congestion and severe congestion. The receiver can then faithfully reflect the ECN signals in the ECN-Echo field. Use of this feedback is left to the sender TCP. TABLE 4.2 PROPOSED ECN AND ECN-ECHO FIELDS ECN ECN Field Meaning ECN- Echo ECN-Echo Field Meaning 00 Not ECN capable 00 Reserved for other use 01 ECN capable and no congestion 01 Echo of no congestion 10 ECN capable and mild congestion 10 Echo of mild congestion 11 ECN capable and severe congestion 11 Echo of severe congestion The routers set congestion signals based on two thresholds, one for mild congestion and another for severe congestion. A router marks the packet according to its current congestion status, but if the packet already carries a higher level mark, it will not change the mark. Since more informational congestion signals are provided, the sender responds to the ECN signals according to the following rules: No Congestion Signal: sender follows existing slow start or congestion avoidance algorithms to increase the window size. Mild Congestion signal: sender decreases its congestion window to (α CWND), where CWND is the current congestion window, if there was no window reduction in the past RTT. Severe Congestion Signal: sender decreases its congestion window to (β CWND), if there was no window reduction in the past RTT. Incase there was a window reduction caused by mild congestion but no window reduction due to severe congestion in the past RTT, then the sender further reduces the congestion window to (β/α CWND). Here, (0 < β < α < 1) The principle in the window reduction is that the sender should reduce its congestion window at most once in an RTT, but a severe congestion signal can override a previous mild congestion signal. Revisiting Table 3.1, we can notice that this approach meets most of the requirements. Since no extra packets are sent there is no traffic overhead, and the feedback is delivered using the IP and TCP headers. Minimal buffering is required because only the unacknowledged packets sent over the wireless link are buffered. Since only a small modification is needed in the existing TCP code, the computing complexity is very small [19]. B. Simulation and Result Analysis In order to find the performance impact, a set of simulations was conducted; TCP performance data for different packet error rates and different packet sizes was collected. The purpose was to compare standard TCP with the proposed scheme. The simulations were performed on a similar simplified network model as used by Raj Jain and Chunlei Liu [18], shown in Figure 4.4. In this model S1 and S2 represent the sources while d1 and d2 represent the destinations. Respective data rates and delays are shown beside each link. The link between d1 and r2 is the wireless link. Fig. 4.3 [18]. Simulation network model The testing traffic is an FTP session from S1 to d1 using TCP as the transport protocol while UDP flow from S2 to d2, used as the background traffic, is produced by an exponential on-off model. The mean burst period and the mean silence period are both set to 100 ms, and the burst data rate is set to 500 kbps. Link layer retransmission is implemented on the wireless link between r2 and d1. Unacknowledged sent packets at the link level are resent within 40ms. Same rate of wireless errors is subjected to these retransmitted packets. The packet error rate of the wireless link is varied to test the performance of the two proposals under different loss scenarios [18]. In our proposal we compare standard TCP with the proposed congestion coherence using three-level ECN. The metrics evaluated to compare the enhancement proposal are congestion losses, retransmissions and throughput. In figures 4.4 and 4.5 we compare the number of congestion losses and the total number of end-to-end retransmissions. By observing figure 4.4 we find that the number of congestion losses of standard TCP is significantly more than congestion coherence using three-level ECN because standard TCP uses packet losses as a congestion control mechanism. With the rise in packet error rate the size of the congestion window decreases again and again due to the presence of wireless losses. Ultimately the congestion window gets so small that no packet is required to be dropped due to congestion.

6 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 11 No: Fig Congestion Losses Fig Number of End-to-End Retransmissions The number of end-to-end retransmissions depends on the number of wireless losses, congestion losses and timeouts, as well as the enhancement method used. Standard TCP retransmits both congestion and wireless losses. Our enhancement proposal not only avoids most of congestion losses, but also has the least count of retransmissions because it awaits the local retransmission of wireless losses. Fig Throughput versus Packets Error Rate Fig Throughput versus Packets Size (bytes) From the figures 4.6 and 4.7 we observe that standard TCP performs fairly well when the packet error rate is very small, but with the rise in packet error rate its performance degradation is rapid. The performance curve confirms the requirement of an enhancement in wireless TCP. Congestion Coherence using three-level ECN improves throughput for all packet error rates in the simulated range. V. CONCLUSION AND FUTURE WORK The goal of wireless TCP enhancements is to identify the wrong assumptions and deficiencies in current networking protocols that are the cause of poor performance, and to reform these networking protocols such that wireless networks can produce high performance and provide highquality network services to mobile users. A. Conclusion In this study, we assessed that the Congestion Coherence with three-level ECN scheme aggravates TCP over wireless links. Using three-level ECN as the congestion feedback mechanism and the congestion coherence in consecutive packets, this scheme avoids majority of end-to-end retransmissions, unnecessary slowdowns and timeouts caused by wireless errors and hence improves the performance of TCP over wireless links. This scheme only requires small changes in the TCP code at the mobile station s side and most of the changes are needed in the base station and the fixed host, assuming that three-level ECN has already been implemented in all network routers. This scheme maintains TCP s end-to-end semantic, and all alterations are in the reach of wireless network service providers. The proposed scheme applies to full duplex traffic and can also work with encrypted traffic. These vital points validate the Congestion Coherence with three-level ECN scheme as a unique wireless TCP enhancement. This scheme explicitly shows the advantages of three-level ECN over the current traditional congestion control mechanisms. Since the proposed scheme highly depends on three-level ECN, we propose it to be used when three-level ECN is widely deployed. A conclusion that can be drawn from this study is that the three-level ECN standard should be deployed widely to improve wireless networks performance. B. Future Work and Scope The proposed enhancement scheme requires the routers in the data path to implement three-level ECN. It requires that all routers implement three-level ECN. Since implementation of three-level ECN in the Internet will be gradual, in future, we need to find alternative ways to implement this scheme when the entire network is not ECN capable. The marking policy for the congested packets has a significant impact on our proposed scheme. When congestion is experienced, it is important to make certain that packets are not dropped without neighboring packets being marked. The randomness in deciding whether to mark a packet helps to desynchronize the TCP congestion windows among flows, but may jeopardize this assurance. A better marking policy should keep a random marking zone to help desynchronize TCP congestion windows among flows, and a deterministic marking zone to ensure congestion coherence. It should also be kept in account, that it is the real queue length that is affiliated to packet losses. Normally lower coherence and low good put is yielded using average queue length, because of its delay in reflecting the real congestion status. REFERENCES [1] Behrouz A. Frouzan, TCP/IP Protocol suite. TATA McGRAW-Hill Edition, 3rd Edition, Fourth reprint 2006.

7 International Journal of Electrical & Computer Sciences IJECS-IJENS Vol: 11 No: [2] J. B. Postel. Transmission Control Protocol. RFC, Information Sciences Institute, Marina del Rey, CA, September RFC-793. [3] M. Allman, S. Floyd, and C. Partridge. RFC 3390: Increasing TCP's initial window, August [4] J. Garcia and A. Brunstrom. Checksum-based loss differentiation. Proceedings 4th IEEE Conference on Mobile and Wireless Communications Networks (MWCN 2002), Stockholm, Sweden, September [5] W. R. Stevens. TCP/IP Illustrated, Volume 1. Addison-Wesley, Reading, MA, Nov [6] H. Balakrishnan, V. N. Padmanabhan, S. Seshan, and R. H. Katz. A comparison of mechanisms for improving TCP performance over wireless links. IEEE/ACM Transactions on Networking, 5(6): , December [7] G. H. Heilmeier. Global begins at home. IEEE Communications Magazine, 30(10):50 56, October [8] R. C aceres and L. Iftode. Improving the performance of reliable transport protocols in mobile computing environments. IEEE Journal on Selected Areas in Communications, 13, June [9] The Working Group for WLAN Standards. IEEE wireless local area networks (webpage). checked [10] M. Taferner and E. Bonek. Wireless Internet Access over GSM and UMTS. Springer-Verlag, [11] Annika Wennstr om, Stefan Alfredsson, and Anna Brunstrom. TCP over Wireless Networks. Karlstad University Studies 2004:21, Karlstad University, Sweden, May [12] M. Kojo, K. E. E. Raatikainen, M. Liljeberg, J. Kiiskinen, and T. O. Alanko. An efficient transport service for slow wireless telephone links. IEEE Journal of Selected Areas in Communications, 15(7): , [13] M. Allman, V. Paxson, and W. Stevens. TCP congestion control. RFC 2581, Apr [14] R. T. Braden. RFC 1122: Requirements for Internet hosts communication layers,oct [15] D. Chiu and R. Jain. Analysis of the increase and decrease algorithms for congestion avoidance in computer networks. Computer Networks and ISDN Systems, [16] K. Ramakrishnan and S. Floyd. A proposal to add explicit congestion notification, [17] S. Floyd. TCP and explicit congestion notification. ACM Computer Communication Review, 24(5):10 23, [18] Chunlei Liu, Raj Jain, Approaches of Wireless TCP Enhancement and A New Proposal Based on Congestion Coherence, Proceedings of the 36th Hawaii International Conference on System Sciences IEEE. [19] Durresi, A., Sridharan, M., Liu, C., Goyal, M. and Jain R. (2001) Traffic management using multilevel explicit congestion notification, Proceedings of the 5th World MultiConference on Systemics, Cybernetics and Informatics SCI 2001, ABR over the Internet, July. [20] R. Jain. A timeout-based congestion control scheme for window flowcontrolled networks. IEEE Journal on Selected Areas in Communications, SAC-4, 7: , [21] S. Floyd. The gentle option in ns, March [22] V. Jacobson. Congestion avoidance and control. ACM Computer Communication Review; Proceedings of the Sigcomm 88 Symposium in Stanford, CA, August, 1988,18, 4: , [23] J. C. Hoe. Improving the start-up behavior of a congestion control sheme for TCP. In Proceedings of the ACM SIGCOMM Conference on Applications, Computer Communications, New York, Aug ACM Press. [24] L. S. Brakmo, S. W. O Malley, and L. L. Peterson. TCP Vegas: New techniques for congestion detection and avoidance. In Proceedings, 1994 SIGCOMM Conference, [25] S. Floyd and V. Jacobson. Random early detection gateways for congestion avoidance. IEEE/ACM Transactions on Networking, 1(4): , Aug [26] S. Banerjee and J. Goteti. Extending tcp for wireless networks. University of Maryland,College Park, [27] N. Vaidya, M. Mehta, C. Perkins, and G. Montenegro. Delayed duplicate acknowledgements: A TCP-unaware approach to improve performance of TCP over wireless. Technical Report, Computer Science Dept., Texas A&M University, [28] S. Biaz, M. Mehta, S. West, and N. H. Vaidya. TCP over wireless networks using multiple acknowledgements. Technical Report , Computer Science, Texas A&M University, [29] H. Balakrishnan and R. Katz. Explicit loss notification and wireless web performance.proc. IEEE Globecom Internet MiniConference, Sydney, Australia, November [30] S. Biaz and N. H. Vaidya. Discriminating congestion losses from wireless losses using inter-arrival times at the receiver. IEEE Symposium ASSET 99, Richardson, TX, USA, March 1999.