TCP protocol has gone through a long history evolving from its primitives such as Reno and NewReno, however, the network topology and technology has been quite different from when the protocol was proposed and this has brought new challenges to it. Theory and experiments show that as per-flow product of bandwidth and latency increases, TCP becomes inefficient and prone to in-stability, regardless of the queuing scheme. What is more, with the large-scale application of some new technology such as Asymmetrical Digital Subscriber Line (ADSL) and wireless network TCP has been proved to be unable to cope with the new environment. To solve the problem, Dina Katabi and Mark Handley et al. has proposed eXplicit Control Protocol (XCP) [1] to make good use of the routers' help to provide more inf...[ Read more ]

TCP protocol has gone through a long history evolving from its primitives such as Reno and NewReno, however, the network topology and technology has been quite different from when the protocol was proposed and this has brought new challenges to it. Theory and experiments show that as per-flow product of bandwidth and latency increases, TCP becomes inefficient and prone to in-stability, regardless of the queuing scheme. What is more, with the large-scale application of some new technology such as Asymmetrical Digital Subscriber Line (ADSL) and wireless network TCP has been proved to be unable to cope with the new environment. To solve the problem, Dina Katabi and Mark Handley et al. has proposed eXplicit Control Protocol (XCP) [1] to make good use of the routers' help to provide more information to the end host thus allowing more flexi-ble and more robust control. Following the idea of XCP, Nandita Dukkipati, Nick McKeown and et al. in Stanford later proposed Rate Control Protocol (RCP) [2] to use the routers as an assistant to tell the end host how many bandwidth it should use to minimize the flow duration time.

The advantages of these two protocols are obvious and with the information feedback from the routers, the end host can make better decision and utilize the valuable network resources more efficiently. However, the ideas they proposed are quite different from the traditional TCP protocol which is end-to-end in nature and are difficult to be deployed. Up to recently concentration has been focused upon the improvement which can still maintain the end-to-end nature of TCP protocol.

This thesis makes observation of the discrete delayed feedback control na-ture of the TCP protocol and proposes an end-to-end TCP protocol based upon Active Bandwidth measurement (AmTCP) with the objective of retaining the nice features of XCP and RCP. Our protocol investigates TCP's discrete delayed feedback control nature and tries to model our algorithm accordingly and thus assumes the best guess of the information which was supposed to be provided by routers in XCP and RCP. The end host tries to form the control process to a round by round procedure and after completion of each round the bottleneck bandwidth will be measured and the end host can then adjust quickly to this new rate. Our objective is that the protocol can quickly converge to the real value of available bandwidth and can adapt quickly to the changes in the available band-width. Our protocol in nature is a rate-based protocol and by measuring out the exact value of the available bandwidth we can smooth the process of sending out packets to avoid false congestion as well as making best utilization of the available bandwidth.

Other important enhancement is initial bandwidth measurement which has been borrowed from techniques in bandwidth measurement areas to obtain a coarse estimate of available bandwidth quickly to get rid of the conventional slow start period. By starting with a relatively high yet reasonable rate, we will be able to improve the end users' experience of the network application especially for those applications heavily biased upon the short flows.

Also, to cater for the more prevalent ADSL technology, which usually employs a lot of asymmetrical links as its carrier, we also add the clock synchronization function to maintain common-clock between the sender and receiver so our band-width measurement will not be affected by the asymmetrical nature of the link or when the forward and reverse links are experiencing different congestion con-ditions.

Fairness issues and robustness against some malicious sources are also dis-cussed in this thesis.