Analysis and design of a reliable wireless transport protocol and fair network mechanisms for the internet

We address two problems concerning transport protocols in this thesis. The first problem involves the operation of the widely used reliable transport protocol namely TCP over wireless links. We develop an analytical methodology for the performance evaluation of a window based transport protocol and make a comparative study of the different TCP algorithms over wireless links with both iid packet losses and correlated packet losses. Towards this, we first specify an analytical model called the packet train model. This model while being simple enough also accurately describes and models the different TCP phases such as slow start, congestion avoidance, fast retransmit and fast recovery. This framework of the packet train model also allows us to model and evaluate a wide class of variations to the basic protocol. Using the packet train model we analyze the performance of the different TCP algorithms over the wireless link. We show that the performance of NewReno is worse than the performance of Tahoe in many situations characterized by correlated packet losses. We also derive conditions, satisfaction of which leads to significant throughput deterioration. To improve the performance of the different TCP algorithms we propose incremental changes to the different TCP algorithms. We then provide conditions on the loss probability values, satisfaction of which ensures that the proposed mechanisms are advantageous in terms of improving the efficiency of the protocol over the wireless links. Finally, we also study the effect of random packet losses, both iid and correlated, on transport protocols with different window adaptation strategies. The second problem that we consider is that of fair bandwidth sharing among adaptive (TCP) and non-adaptive (i.e. CBR-UDP) flows at an Internet gateway. An algorithm that drops packet preventively, in an attempt to actively penalize the non-adaptive traffic that attempts to “steal” buffer space, and therefore bandwidth from the adaptive traffic flows, is presented. The algorithm maintains minimal flow state information and is therefore scalable. The performance of the algorithm is compared with other gateway algorithms and it is shown that, in the presence of non-adaptive traffic, it achieves a more balanced bandwidth allocation among the different flows.