Paper Review : A Comparison of Mechanisms for Improving TCP Performance over Wireless Links


Reviewer : Seh Leng Lim


This paper is an attempt to compare various feasible mechanisms to improve TCP performance over wireless links.


The main contribution of the paper is its proposal that a reliable link-layer protocol that is TCP-aware provides very good performance, and that it is possible to achieve good performance without splitting the end-to-end connection at the base station. It also recommends that selective acknowledgements (such as SACK or SMART) and explicit loss notifications (ELN)  result in significant performance improvements.


The key main ideas expounded are:

     (a) Hiding noncongestion related losses from TCP

(b)    Making sender aware of the existence of wireless hops so as to avoid invoking congestion control algorithms when noncongestion related losses occur

(c)    Coexisting of wireless-aware transport protocol with link layer schemes


The authors classified the various schemes to alleviate the effects of noncongestion-related losses on TCP performance over wireless links as end-to-end proposals, split-connection proposals, and link-layer proposals. The end-to-end protocols either use selective acknowledgements or explicit loss notification. On the other hand, split-connection hides the wireless link from the sender by terminating the TCP connection at the base station. Link layer solutions attempt to hide link-related losses from the TCP sender by using local retransmissions and perhaps forward error correction.


Link layer protocols (such as LL in the case of the paper) suffer from poor TCP performance because of competing retransmissions caused by an incompatible setting of timers at the two layers (between link layer and TCP) and unnecessary invocations of the TCP fast retransmission mechanism due to out of order delivery of data. LL-TCP-aware protocol which makes use of TCP semantics to prevent duplicate acknowledgements caused by wireless losses to reach the sender and locally retransmits packets achieves better performance than LL.


Using the original TCP reno over wireless links results in low throughput due to the large number of timeouts. The resulting average window size during the transfer is small, preventing the “data pipe” from being kept full and reducing the effectiveness of the fast transmission mechanism.


E2E-NEWRENO, E2E-ELN, E2E-SMART and E2E-IETF-SACK each use new TCP options to identify and retransmit lost packets more accurately, and recover from multiple losses in a single transmission window without timing out. In fact, having both ELN and SACK will result in good performance.


For split connection protocols, SPLIT which uses TCP Reno over the wireless connection, and SPLIT-SMART which uses SMART-based selective acknowledgement scheme are used. Both perform worst than that of a well-tuned, TCP-aware link layer protocol. This demonstrates that end-to-end connection need not be split at the base station to achieve good performance.


I think that the paper has a significant contribution (rating of 4) to the study of how to improve TCP performance in the wireless networks. Through their experimental results, the authors conclude that :


(a)    a reliable link-layer protocol with some knowledge of TCP results in very good performance

(b)    it is possible to achieve good performance without splitting the end-to-end connection at the base station

(c)    the use of selective acknowledgements and explicit loss notifications result in significant performance improvements


However, the experiments which make use of simulated exponential bit-error rates as well as a test configuration of a sender and a receiver may be a little simplistic. Therefore, it remains to be seen whether the author’s conclusions still hold in a realistic setting of day-to-day use. As the authors studied burst errors only in terms of burst length of 2, 4 and 6 packets, they also conceded that this was simplistic, and suggested that a temporal-loss model based on average lengths of fades and other causes of wireless losses  may be more realistic. They also suggested studying the various protocols under various network topologies.


Researchers and builders who build wireless applications will have a better appreciations from this paper of the difficulties faced in using TCP over wireless links.