Modeling web/TCP transfer latency

Date
2004
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
This thesis proposes a new mathematical model to predict TCP transfer latency for shortlived flows with random packet losses, which are representative of TCP flows carrying Web traffic. Unlike earlier stochastic models, the new model characterizes a data transfer as alternating cycles, with TCP state information carried over from one cycle to the next. Model validation through simulation experiments indicates that the new model's predictions match the simulation results better than earlier models [4][5][40], and surprisingly, the new model fits long-lived TCP flows as well. The proposed model is then extended to Partial CATNIP TCP. CATNIP TCP [49] classifies packets in a flow as high and low priority based on the observation that not all TCP packet losses have the same effect on the data transfer time. Partial CATNIP TCP is different from CATNIP TCP in that it does not consider as high-priority packets sent when the TCP congestion window size is small. Simulation experiments show that the differences between Partial CATNIP TCP and CATNIP TCP are minimal when the packet loss probability is less than 10%, for short-lived flows and long-lived flows. The Partial CATNIP TCP model is validated through simulation experiments. The results show that the Partial CATNIP TCP model fits the simulation closely. In addition, performance comparisons between Partial CATNIP TCP and TCP Reno demonstrate that for short-lived flows, Partial CATNIP TCP is 10% (or more) faster than TCP Reno in most cases. For long-lived flows, Partial CATNIP TCP overall is 7% (or more) faster than TCP Reno. Therefore, CATNIP TCP is a suitable approach to improve TCP performance.
Description
Bibliography: p. 92-98
Some pages are in colour.
Keywords
Citation
Li, Y. P. (2004). Modeling web/TCP transfer latency (Master's thesis, University of Calgary, Calgary, Canada). Retrieved from https://prism.ucalgary.ca. doi:10.11575/PRISM/22727
Collections