We analyze the RTO algorithm standardized for TCP and SCTP [RFC2988] that is widely deployed in the Internet, referred to as the RFC2988-RTO in this paper. We briefly demonstrate three well-known problems of the RIFC2988-RTO. We then develop a new RTO algorithm called the Peak-Hopper-RTO (PH-RTO) that eliminates the mentioned problems. Through extensive simulations in ns2, we evaluate the effectiveness of both RTO algorithms. The key advantage of the PH-RTO over the RFC2988-RTO is its predictability. Although the RFC2988-RTO often exhibits low loss detection times, it also exhibits frequent "RTO outliers" that can lead to exceptionally long loss detection times. The loss detection times of the PH-RTO are much more closely spread around the mean. Finally, our results show that the PH-RTO is much more robust to sudden delay spikes that are particularly common in wireless networks.
