An enhanced slow-start mechanism of TCP Vegas

• Main Author: 張凱翔
• Other Authors: 詹益禎
• Format: Others
• Language: zh-TW
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/85942142156119792794

碩士 === 國立彰化師範大學 === 資訊工程學系 === 95 === With the flourishing development of network technology, the network bandwidth has been greatly increased. If the tradtiotional TCP (transmission control protocol) is adopted in such progressing network, TCP will eventually become the performance bottleneck itself. Therefore, how to enhance the efficiency of TCP in high speed networks becomes the hot topic among the researchers in recent years. In this thesis the slow-start mechanism of TCP Vegas is investigated and, subsequencely, an enhanced slow-start scheme is proposed that try to improve the performance of TCP Vegas in high speed networks. TCP Vegas calculates the number of packets queued in the bottleneck buffer ( ) as a sign of congestion. When >γ, Vegas leaves its slow-start phase and enters the congestion avoidance phase. However, the calculation of may be affected by the burstiness problem that is inheret in the TCP transmission. As a result, Vegas may leave the slow-start phase too early and performance suffers. This issue is more serious when Vegas works in high speed networks. Two techniques are presented to improve the slow-start mechanism of TCP Vegas. First, to alleviate the burstiness transmission of Vegas in its slow-start phase, half of congestion window is increased every round-trip time (RTT) instead of doubling the congestion window every other RTT. Second, a new bandwidth estimator that can calculate the available bandwidth of a connection. The calculated bandwidth is used to dynamically configure the value of γ. Thus, the setting of γcould adapt to the variation of networks, and therefore the congestion window size would may be a proper value while Vegas leaves the slow-start phase. Simulation results show that the enhanced slow-start mechanism and the original one have similar behaviors in low bandwidth networks. However, the proposed mechanism demonstrates significant improvement when connection works in high speed networks. The throughput improvement may reach about 20%~30%.