Comparison of weighted-average and median filters for wireless retransmission timeout estimation

Estimation of retransmission timeout (RTO) in TCP networks is important in order to maximize throughput. If the RTO is too small, packets will be unnecessarily declared lost, thus resulting in a retransmission. If RTO is too large, applications will wait too long before declaring a packet to be lost, thus reducing throughput. The conventional method due to Karn and Jacobson for estimating RTO is using a smoothing filter [1], [2], and this was later refined to utilize mean deviation measurements [1], [2]. In wireless links, the delay can often be impulsive, and weighted-average filters perform poorly. This paper presents results for mobile wireless transmission, and investigates the use of a median filter for timeout estimation in the presence of competing traffic flows. We demonstrate the applicability of this approach to estimating TCP round-trip times, and UDP packet jitter.