Towards Modeling Effect of Packet-level TCP Dynamics on Internet Operations and Management

Published

April 16, 2012

Author(s)

Vladimir V. Marbukh

Abstract

This paper reports on work in progress on modeling of the effect of packet-level TCP congestion avoidance dynamics on flow-level Internet performance and management including buffer sizing under arriving/departing flows. The proposed model accounts for the packet-level TCP burstiness due to congestion avoidance dynamics by incorporating the entire effective bandwidth of TCP-controlled traffic as opposed to the conventional model which accounts for the average TCP rate only. In particular, we demonstrate that conventional Markov model of Internet performance is applicable only in a case of sufficiently heavy average load corresponding to long flow-level queues and file transfer times. In a case of lighter average load the conventional Markov performance model significantly underestimates flow-level queues and file transfer times, and surprisingly a simpler mean-field performance model may be more accurate. While increasing buffer sizes causes proportional increase in the worst-case packet-level delays, the positive effect on the average capacity utilization and flow-level performance may be muted due to increase in the TCP burstiness. These conclusions may explain and quantify observed self-similarity of TCP traffic, and have important implications for the Internet design, operations, and management.

Proceedings Title

2012 IEEE/IFIP Network Operations and Management Symposium

Conference Dates

April 16-20, 2012

Conference Location

Maui, HI

Pub Type

Conferences

Download Paper

Local Download

Keywords

packet-level TCP dynamics, flow-level Internet performance, buffer sizing

Citation

Marbukh, V. (2012), Towards Modeling Effect of Packet-level TCP Dynamics on Internet Operations and Management, 2012 IEEE/IFIP Network Operations and Management Symposium, Maui, HI, [online], https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=910756 (Accessed April 18, 2024)