The performance evaluation and comparison of TCP-based high-speed transport protocols

With the development of network technology, the high-speed networks with 1Gbps, 10Gbps and even higher bandwidth have emerging. In the traditional IP networks, TCP protocol performs well, However, in fast long distance networks (FLDnet), TCP's slow start and its Additive Increase Multiplicative Decrease (AIMD) congestion control are too conservative to make full use of the abundant bandwidth resource. This paper mainly evaluates the TCP-based improved protocols in the Linux 2.6.18 kernel by establishing point-to-point and multipoint-to-point test beds. We compare the effect of RTT and packet loss rate to the transmission performance of the TCP-based protocols. We also evaluate the intra-protocol fairness, TCP-friendliness and inter-RTT fairness of these protocols. The results show that: HSTCP, HTCP, STCP, BIC TCP, and CUBIC TCP have relatively better transmission performance. When the number of the parallel flows is less than four, the protocols of TCP Reno, HSTCP, HTCP, and STCP have good intra-protocol fairness. TCP Reno, HSTCP, and HTCP have good TCP-friendliness and inter-RTT fairness, but the overall throughput is not high enough. The TCP-friendliness and inter-RTT fairness of STCP, BIC TCP and CUBIC TCP are not very good, but using these protocols can improve network transmission performance.