Stabilized Vegas

TCP Vegas was introduced in 1994 [5] as an alternative to TCP Reno. Unlike Reno (or its variants such as NewReno and SACK), that uses packet loss as a measure of congestion, Vegas uses queueing delay as a measure of congestion [15, 18]. Vegas introduces a new congestion avoidance mechanism that corrects the oscillatory behavior of AIMD (Additive Increase Multiplicative Decrease). While the AIMD algorithm induces loss to learn the available network capacity, a Vegas source adjusts its sending rate to keep a small number of packets buffered in the routers along the path. Provided there is enough buffering, a network of Vegas sources will stabilize around a proportionally fair equilibrium and packet loss will be eliminated; see [15] for details. In this paper, we study the stability of this equilibrium in the presence of network delay, motivated by two lines of recent research.