Improving wireless TCP/IP performance using the median filter algorithm

Abstract

The estimation of Retransmission Timeout (RTO) in Transmission Control Protocol (TCP) affects the throughput of the transmission link. If RTO is just a little larger than Round Trip Time (RTT), retransmissions will occur too often, and this increases congestion in the transmission link. If RTO is much larger than RTT, the response to retransmit when a packet is lost will be too slow, and this will decrease the throughput in the transmission link (Comer 2006a). Currently, the Jacobson/Karels Algorithm is widely used for the estimation of RTO in TCP implementations. The algorithm uses an Exponential Weighted Moving Average (EWMA) filter to estimate RTT and then determines the RTO from this. The EWMA filter is good if the RTT follows a Gaussian distribution. In reality, traffic in the Internet is bursty and tends to follow a heavy-tailed distribution. Using an EWMA approach to estimate heavy-tailed distribution is inadequate. The median filter has been recognized as a useful non-linear filter due to its edge preserving and impulse suppressing characteristics, so it is effective in removing impulsive noise (Nodes & Gallagher Jr. 1982). The median filter has been applied to many areas of signal processing, particularly in image processing to remove positive and negative impulsive noise. Thus, it can perform well for heavy-tailed distributions. In this project, the median filter is applied to estimate the RTT over TCP links, under bursty traffic conditions. Experiments and simulations are conducted to determine if the median filter performs better than the EWMA filter. In the experiments, consistent RTT and a small RTO are obtained, which are desirable factors for high connection throughput. In the simulations, the median filter not only delivers higher throughput, but also drops fewer packets during transmission than EMWA filter does. Both the experiments and simulations show that the median filter outperforms the EWMA filter under bursty traffic conditions.