High speed data wireless networks in multipath environments suffer channel impairment from many sources such as thermal noise, path loss, shadowing, and fading. In particular, short-term fading caused by mobility imposes irreducible error floor bounds on system performance. We study the effect of fading on the performance of the widely used TCP/UDP protocol, and investigate how to improve TCP performance over fading channels. Our solutions target upcoming mobile wireless systems such as IEEE 802.16e wireless MANs "Metropolitan Area Networks" where adaptive modulation is enabled and the underlying medium access scheme is On-Demand Time Division Multiple Access "On-Demand TDMA". Adaptive modulation is used in the new generation of wireless systems to increase the system throughput and significantly improve spectral effciency by matching parameters of the physical layer to the time-varying fading channels. Most high-rate applications for such wireless systems rely on the reliable service provided by TCP protocol. The effect of adaptive modulation on TCP throughput is investigated. A semi-Markov chain model for TCP congestion/flow control behavior and a multi-state Markov chain model for Rayleigh fading channels are used together to derive the steady state throughput of TCP Tahoe and Reno. The theoretical prediction based on our analysis is consistent with simulation results using the network simulator NS2. The analytical and simulation results triggered the idea of cross-layer TCP protocol design for single-user scenarios. The fading parameters of wireless channels detected in the physical layer can be used to dynamically tune the parameters "such as packet length and advertised receiver window size" of the TCP protocol in the transport layer so that TCP throughput is improved. For multi-user scenarios, we study how multi-user diversity can be used to improve th.