Statistical analysis of TCP's retransmission timeout algorithm
IEEE/ACM Transactions on Networking (TON)
| Hi-index | 0.01 |
Congestion control is important to the performance of the Internet and this dissertation focuses on congestion control mechanisms closely related to signal processing, including TCP's retransmission timeout (RTO) algorithm, and the random early detection (RED) algorithm. We first conduct the first rigorous analysis of TCP's RTO algorithm. The RTO algorithm can be considered as a filtering problem where the inputs are the round trip time (RTT) measurements, and the outputs are RTO values. We construct a statistical model of the RTT process based on an experimental result that the RTTs of a given Internet path are approximately Gamma-distributed, and an assumption that the temporal correlation of the RTTs is weak. We then evaluate the current RTO algorithm against two established performance measures: the mean RTO and the premature timeout probability , and derive a closed form expression for the first measure and an expression (to be numerically evaluated) for the second. Our analysis results are validated through simulations and strengthen some observations reported in the literature. We then consider improving the performance of the current RTO algorithm (Jacobson's algorithm), which is linear and is susceptible to RTT outliers. In particular, we propose a robust alternative RTO algorithm based on Recursive Weighted Median (RWM) filtering. The RWM-based RTO algorithm utilizes the robust characteristics of RWM filters to form possibly more reliable RTO estimates, thereby possibly leading to better RTO performance. Simulations show that for impulsive RTT behavior the proposed RTO algorithm outperforms Jacobson's algorithm, and for other RTT behavior they result in comparable performance. We finally investigate Internet congestion control by looking into the interaction between TCP Reno and the random early detection (RED) algorithm. Because of its linearity, the RED algorithm cannot accommodate the need to smooth out transient increases in the queue size, and the need to track sudden decreases in the queue size at the same time. To overcome this problem, we design an adaptive rank-based nonlinear filter—the adaptive median filter for the estimation of the queue size. We then analyze the interaction between the proposed Median RED algorithm and TCP Reno in a network system, and show through analysis and simulations that the Median RED algorithm can yield better performance than the RED algorithm.