It is well known that the behaviors of wireless and wireline networks are different as far as packet losses are concerned. Wireline networks experience few packet losses, caused mainly by network congestion. In contrast, wireless networks are noisy, characterized by bursty and high channel error rates. As a consequence, the transmission control protocol (TCP) algorithms developed for wireline networks are not quite adequate for wireless data transmission. This paper analyzes the performance of certain TCP algorithms (OldTahoe, Tahoe, NewReno, and SACK) over a wireless channel, with correlated packet losses.
When a packet loss is detected by a network using TCP, by timeout or duplicate acks, the congestion window is reduced, and a recovery algorithm initiated. While this approach is not the best fit for a wireless channel, and although alternative mechanisms have been proposed, and even experimentally implemented, wireless networks still use one of the well-known TCP versions. Comparing them, and choosing the best suited one, makes sense. While most previous work used simulations, with their inherent limitations, very little research has examined an analytical model of the issue. The authors of this paper have developed such a model. The paper describes it, presenting its limitations and its uses very clearly.
The authors analyze the performance of the different TCP versions, checking their sensitiveness to bursty conditions, to the granularity of the timeout timer, and to other factors. From a practical point of view, the selective acknowledgement TCP was found to be superior in most cases.
This topic will certainly generate future papers addressing more general and complex models, like multiflow TCP and different kinds of packet loss. This paper will give the interested reader a deeper phenomenological insight into TCP throughput in wireless channels, while also providing well-chosen references.