The purpose of this paper is to describe another implementation of TCP (Vegas) that has 37 to 71% higher throughput on the Internet than the current distribution (Reno) of BSD Unix.
The main contribution of this paper is to concretely describe a new implementation of TCP.
1) The Vegas implementation does not involve any changes to the TCP
specification, as all of the changes are confined to the sending side.
2) The improvement in throughput is not achieved by an aggressive retransmission strategy that effectively steals bandwidth away from TCP connections that use the Reno implementation. Rather, it is achieved by a more efficient use of the available bandwidth.
3) The three new techniques implemented in Vegas are (1) a new retransmission mechanism, (2) a congestion avoidance mechanism, and (3) a modified slow-start mechanism.
I would rate this paper as a 5 because it significantly improves the throughput on the Internet by providing a better implementation of TCP without requiring any modifications to the protocol itself.
By presenting both simulated and actual performance results, they make an effective argument to upgrade to the Vegas implementation. I thought it was particularly interesting that "Reno connections do slightly better when the background traffic is running on top of Vegas as compared to when the traffic is running on top of Reno." (In fact, the authors seem to emphasize this point; a choice that I believe strengthens their argument).
Given the multitude of tests (both actual and simulated) run both the authors, as well as the extended discussion at the end of the paper, I found very few limitations in the authors approach. The only potential limitation I saw was that the authors presented a weak defense for Vegas as being a fair algorithm: "we conclude that Vegas is no less fair than Reno."
One lesson researchers should take away from this work is that it is often possible to improve upon existing implementations of a given protocol without having to modify the protocol itself.