This thesis presents research on transport layer behavior in wireless networks. As the Internet is expanding its reach to include mobile devices, it has become apparent that some of the original design assumptions for the dominant transport protocol, TCP, are approaching their limits. A key feature of TCP is the congestion control algorithm, constructed with the assumption that packet loss is normally very low, and that packet loss therefore is a sign of network congestion. This holds true for wired networks, but for mobile wireless networks non-congestion related packet loss may appear. The varying signal power inherent with mobility and handover between base-stations are two example causes of such packet loss. This thesis provides an overview of the challenges for TCP in wireless networks together with a compilation of a number of suggested TCP optimizations for these environments. A TCP modification called TCP-L is proposed. It allows an application to increase its performance, in environments where residual bit errors normally give a degraded throughput, by making a reliability tradeoff. The performance of TCP-L is experimentally evaluated with an implementation in the Linux kernel. The transport layer performance in a 4G scenario is also experimentally investigated, focusing on the impact of the link layer design and its parameterization. Further, for emulation-based protocol evaluations, controlled packet loss and bit error generation is shown to be an important aspect.

In this thesis, the performance of the wireless Internet is studied with a focus on TCP over GSM/GPRS. To this end, we have developed a GSM/GPRS test bed for TCP measurements. We present measurements conducted in the test bed with TCP over both GSM and GPRS. The GSM measurements show that TCP interacts efficiently with the radio link layer in GSM, also under very poor radio conditions. The only cases when some inefficiency is detected occur due to sudden variations in delay. In GPRS, delay is, among other things, caused by buffering in intermediate nodes and by dynamic resource sharing. The GPRS measurements indicate that if the buffers in intermediate nodes are reduced, then the delay can be decreased significantly with almost retained throughput. Dynamic resource sharing is often configured to give preemptive priority to voice calls, which implies that all GPRS resources may be lost. Measurements with preemptive priority indicate that TCP performance is degraded more than necessary due to inefficient buffer management. The situation would improve if data was buffered during preemption and if the data was transmitted immediately as GPRS resources become available again