Date of Award

Spring 2005

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computer Engineering - (Ph.D.)

Department

Electrical and Computer Engineering

First Advisor

Nirwan Ansari

Second Advisor

Sirin Tekinay

Third Advisor

MengChu Zhou

Fourth Advisor

Yun Q. Shi

Fifth Advisor

Cristian Borcea

Abstract

The transmission control protocol (TCP) exhibits poor performance when used in error-prone wireless networks. Remedy to this problem has been an active research area. However, a widely accepted and adopted solution is yet to emerge. Difficulties of an acceptable solution lie in the areas of compatibility, scalability, computational complexity and the involvement of intermediate routers and switches.

This dissertation rexriews the current start-of-the-art solutions to TCP performance enhancement, and pursues an end-to-end solution framework to the problem. The most noticeable cause of the performance degradation of TCP in wireless networks is the higher packet loss rate as compared to that in traditional wired networks. Packet loss type differentiation has been the focus of many proposed TCP performance enhancement schemes. Studies conduced by this dissertation research suggest that besides the standard TCP’s inability of discriminating congestion packet losses from losses related to wireless link errors, the standard TCP’s additive increase and multiplicative decrease (AIMD) congestion control algorithm itself needs to be redesigned to achieve better performance in wireless, and particularly, high-speed wireless networks. This dissertation proposes a simple, efficient, and effective end-to-end solution framework that enhances TCP’s performance through techniques of adaptive congestion control and active queue management. By end-to-end, it means a solution with no requirement of routers being “wireless-aware” or “wireless-specific”.

TCP-Jersey has been introduced as an implementation of the proposed solution framework, and its performance metrics have been evaluated through extensive simulations. TCP-Jersey consists of an adaptive congestion control algorithm at the source by means of the source’s achievable rate estimation (ARE) —an adaptive filter of packet inter-arrival times, a congestion indication algorithm at the links (i.e., AQM) by means of packet marking, and a effective loss differentiation algorithm at the source by careful examination of the congestion marks carried by the duplicate acknowledgment packets (DUPACK).

Several improvements to the proposed TCP-Jersey have been investigated, including a more robust ARE algorithm, a less computationally intensive threshold marking algorithm as the AQM link algorithm, a more stable congestion indication function based on virtual capacity at the link, and performance results have been presented and analyzed via extensive simulations of various network configurations. Stability analysis of the proposed ARE-based additive increase and adaptive decrease (AJAD) congestion control algorithm has been conducted and the analytical results have been verified by simulations. Performance of TCP-Jersey has been compared to that of a “perfect”, but not practical, TCP scheme, and encouraging results have been observed. Finally the framework of the TCP-Jersey’s source algorithm has been extended and generalized for rate-based congestion control, as opposed to TCP’s window-based congestion control, to provide a design platform for applications, such as real-time multimedia, that do not use TCP as transport protocol yet do need to control network congestion as well as combat packet losses in wireless networks.

In conclusion, the framework architecture presented in this dissertation that combines the adaptive congestion control and active queue management in solving the TCP performance degradation problem in wireless networks has been shown as a promising answer to the problem due to its simplistic design philosophy complete compatibility with the current TCP/IP and AQM practice, end-to-end architecture for scalability, and the high effectiveness and low computational overhead. The proposed implementation of the solution framework, namely TCP-Jersey is a modification of the standard TCP protocol rather than a completely new design of the transport protocol. It is an end-to-end approach to address the performance degradation problem since it does not require split mode connection establishment and maintenance using special wireless-aware software agents at the routers. The proposed solution also differs from other solutions that rely on the link layer error notifications for packet loss differentiation.

The proposed solution is also unique among other proposed end-to-end solutions in that it differentiates packet losses attributed to wireless link errors from congestion induced packet losses directly from the explicit congestion indication marks in the DUPACK packets, rather than inferring the loss type based on packet delay or delay jitter as in many other proposed solutions; nor by undergoing a computationally expensive off-line training of a classification model (e.g., HMM), or a Bayesian estimation/detection process that requires estimations of a priori loss probability distributions of different loss types.

The proposed solution is also scalable and fully compatible to the current practice in Internet congestion control and queue management, but with an additional function of loss type differentiation that effectively enhances TCP’s performance over error-prone wireless networks.

Limitations of the proposed solution architecture and areas for future researches are also addressed.

Share

COinS