TCP over military networks.

The major issues with existing military communications are the requirement to improve connectivity between systems. At present there are many bespoke systems used by co-operating nations that are unable to share information because of inter-connectivity problems. The solution requires the use of a common standard that the different nations are willing to implement and use. One network technology that is able to address these problems is the Internet protocol suite. Tactical communications rely heavily on radio links to provide connectivity across the network. Radio links are inherently less reliable than fixed. The aim of this thesis is to study degrading effects at the radio link layer and see how this affects the performance of TCP/IP. The study of the radio link effects concentrated on the distribution of errors (error patterns). The results from radios were compared against standard models and revealed a discrepancy between radios and the models. The military radio showed characteristics that were very different from any of the standard models hence a new burst error model was developed that could mimic these error patterns. The next stage of the study was to observe how these burst errors affect the data link layer protocols. A hypothesis was proposed that burst errors would improve the performance of unprotected packet but adversely affect FEC. This was proven to be true, but the effect is only really noticeable at the point where the throughput collapses. The results also showed that the worst case to study for unprotected packets was random errors not burst errors. The IP layer has no recovery mechanism, it is Transmission Control Protocol (TCP) that provides the reliability. TCP contains several mechanisms to handle lost packet detection, retransmission and congestion management. The performance of each of the algorithms was tested in the presence of errors to determine which combination produced the highest throughput. A number of network restrictions where also detailed, the round trip time delay should be less than 9 seconds and the residual bit error rate should be better than 1 in 10.;Various ways of separating congestioncontrol and corruption were considered to try and improve the performance of TCP in the presence of errors. These mechanisms were TCP Vegas, a modified version of Vegas and Packet Pair measurement. Both the Modified Vegas and Packet Pair Control produced throughputs that were better than standard TCP in the presence of errors. The Modified Vegas produced very good results under congestion, again better than standard TCP. The other approach that was considered for improving the performance of TCP was to change the links to reduce the errors that effect the performance rather than changing TCP to handle errors. The use of ARQ greatly improved the throughput. The use of ARQ increased the variance in the round trip time which has a major impact on the time delay sensitive algorithms such as Vegas and Modified Vegas. It is recommended that a combination of Modified Vegas and ARQ is not used. From the results it is clear that there could be interaction effects between different protocol layers. Burst bit error may or may not produce burst packet losses. Pulse errors can have a dramatic impact on the throughput of TCP. The time delay variance introduced by ARQ can greatly effect time sensitive congestion control mechanisms such as Vegas. Hence with any system it is necessary to consider all layers not just the performance of a single layer under some arbitrary condition