Experiment setup

Split-connection detection

TCP sender throughput comparison

Papers

· Inference and Evaluation of Split-Connection Approaches in Cellular Data Networks
Wei Wei,Chun Zhang, Hui Zang, Jim Kurose, Don Towsley
Passive and Active Measurement Conference 2006, Adelaide, Australia, March 30-31, 2006.

 

Wei Wei's Home Page

MICN

Measurement and Inference of Cellular Networks

Overview

 

Compared with a wireless LAN, a cellular data network provides much wider coverage but at a considerably lower bandwidth.  TCP, the most widely used transport protocol in the Internet, performs poorly in a cellular data network due to varying wireless channel conditions. Numerous mechanisms have been proposed for improving TCP performance over wireless links, including those in wireless cellular networks. One important class of techniques is split-connection techniques, which split a TCP connection in order to isolate wireless related issues from the wired network. We ask the following question: Have split-connection techniques been deployed in commercial cellular data networks? If so, what are the performance gains from these techniques?

 

To answer this question, we develop passive and active methods to detect whether a cellular provider implements split-connection techniques. The essence of both methods is detection by contradiction. We first assume that the TCP connection is not split. If the principle ``Time cannot go backwards'' is violated under this assumption, we conclude that the TCP connection is split. Through end-to-end measurements over commercial cellular networks operated by three different providers (including two CDMA2000 networks and one GPRS network), we find that all three providers selectively implement split-connection techniques for certain applications (e.g., HTTP) and the implementations differ from provider to provider. We further design experiments to investigate performance gains from split-connection techniques and demonstrate that the performance gain can be significant in some cases. Moreover, we explain the surprising relationship between performance gains and TCP flow sizes using analysis. We also explore practical issues in implementing split-connection techniques, e.g., whether the TCP connections are terminated gracefully when a connection is aborted.

 

Traces:

Traces on desktops are collected using tcpdump while traces on laptops are collected using WinDump.

FTP Traces:

HTTP Traces:

TCP Traces: