Congestion is a major Internet problem. TCP has evolved in attemptting to deal with this and now employs congestion control algorithms that effectively limit the bandwidth available to any one connection. However, as TCP is not the only transport protocol used on the Internet, the growth of non-TCP traffic has lead to the formulation of TCP-Friendly formulae. It is recommended that all best effort traffic conform to these formulae, and they act as a benchmark against which the success of any congestion control algorithms can be judged. Unfortunately, key assumptions behind the formulation of TCP's algorithms - low bandwidth, and bulk data transfers - are often no longer valid. In this paper it is shown that the combination of increasing bandwidth and short web transfers means that a significant amount of TCP traffic fails to obtain the bandwidth implied by the TCP-friendly formulae. An alternative, measurement-based approach that predicts a fairer starting window size for a connection is presented and evaluated. Information about the characteristics of particular network paths is dynamically maintained and used to suggest a fairer starting window size for new connections. Minor modifications to TCP allow these suggestions to be used to set start-up control variables, and thereby strengthen the negative correlation between the bandwidth used by a connection and the level of congestion on its network path. The measurement-based approach is realised through the use of a Location Information Server (LIS). The LIS performs centralised passive monitoring of transport headers in order to derive network level path information. Location Information Packets (LIP) are used to communicate suggested start-up variables to local hosts. The design, implementation and evaluation of an LIS, LIP and participating host's TCP software are presented.