The purpose of this paper is to compare several schemes designed to improve the performance of TCP in wireless and other lossy link networks.
The main contribution of this paper is to compare the effectiveness of various TCP-aware schemes for use in a wireless environment. This is important because TCP assumes congestion in the network is the primary cause for packet loss and unusual delays. When packets are lost in networks for reasons other than congestion (such as sporadic high bit-error rates and intermittent connectivity due to handoffs), congestion control measures result in an unneccessary reduction in end-to-end throughput (ie: suboptimal performance).
(1) There are two different approaches to improving TCP performance in
a wireless (or any lossy) system: (a) hide any noncongestion-related
losses from the TCP sender and (b) make the sender aware of the existence
of wireless hops and that some packet losses are not due to congestion.
(2) The schemes are classified into three basic groups: (a) end-to-end proposals, split-connection proposals, and link-layer proposals.
(3) This study lead to the following conclusions: (1) shielding the sender from duplicate acks arising from wireless losses gives a 10-30% higher throughput, (2) splitting the end-to-end connection is _not_ a requirement for good performance, (3) selective ack schemes are very useful in the presence of lossy links, especially when losses occur in bursts, and (4) end-to-end schemes are promising since significant performance gains can be achieved without any extensive support from intermediate nodes in the network.
I would rate this paper as a 3 because, while it didn't propose any new algorithms, it did provide a careful overview and evaluation of a variety of techniques to solve an existing problem.
I feel that by controlling their experiment in such a way to attempt to ensure that losses are only due to wireless errors (and not congesion), the authors are really isolating the problem at its source. In addition, the authors have implemented an exponentially distributed bit-error model over the lossly link. In order to ensure a fair basis for comparison, none of the protocol implementations introduce any additional data copying at intermediate points from sender to receiver.
The most important limitation of this approach is that I would have liked to have seen tests performed in conditions where there is a large variation in connection round-trip times. In general, the approach seems solid.
One lesson researchers should take away from this work is that advancement in new technologies (such as wireless and other lossy networks) may not always follow the same assumptions as previous technologies (in that packet loss isn't always due to congestion). These potentially overlooked areas can lead to some interesting problems and need for future research.