We focus in this paper on the undesirable phenomenon of frequent losses during slow-start when TCP operates in large bandwidth-delay product networks such as those including geostationary satellite links. This phenomenon, already identified in [14,7], results in a degradation in TCP throughput due to a double consecutive slow-start phases per TCP cycle. Given the high cost and the scarcity of satellite links, it is of importance to find solutions to get rid of these losses. We propose two simple modifications to TCP algorithms and illustrate their performance by mathematical analysis and simulations. First, we decrease the slow-start threshold (with respect to current TCP) to less than the window size at which these losses occur, and second, we space the transmission of packets sent during slow-start.