posted on 2022-03-28, 02:14authored byPascal Donfack
Nowadays Transmission Control Protocol/Internet Protocol (TCP/IP) is known to be the most reliable network communication protocol capable of handling retransmission, packet loss, congestion control and more. In TCP/IP communication, packet loss and congestion are likely to have an impact on bandwidth. Performance is affected even at low packet loss rates, so with an enlarged rate of packet loss, a radical drop in bandwidth efficiency occurs. To identify the source of this unpredictable network performance, a thorough examination of TCP/IP traffic was conducted.
This thesis studied the behaviour of the main protocols involved in the 4 layers of the TCP/IP stack. This project, which was primarily concerned with layers 3 and 4 of the TCP/IP stack protocol, carried out a comparison between the performance of two different congestion and control algorithms (Cubic and Reno). Several experimental tests were conducted to determine when the connection experienced data loss and ACK loss, with the results plotted on individual graphs showing packet behaviour.
Initially two different TCP congestion control algorithms were used to observe their influence on bandwidth rate. Subsequently, the TCP variables - advanced window scaling and window scaling - were changed to observe their role in obtaining acceptable bandwidth rate with respect to packet drop and the retransmission rate.
The study revealed a significant reduction in performance during packet loss. Surprisingly, the results showed that the congestion and control algorithms, Cubic and Reno, led to the same outcome when the link was experiencing packet loss. However, ACK loss did not significantly affect performance, and up to 40% loss in ACKs could be tolerated with almost no reduction in performance.
While TCP still functions adequately when experiencing single segment loss, the challenge is to handle multiple packet loss. Further research, therefore, might investigate new algorithms for multiple packet loss in the network.