TCP mechanisms for Diff-Serv Architecture
Work on Diff-Serv has demonstrated that it is possible to
create differentiations in throughput among TCP connections
during periods of network congestions. However, the effectiveness
of such schemes is limited by the impreciseness and biases in
TCP's window-based congestion control algorithm. More precisely,
TCP's window-opens mechanism has an intrinsic bias agains long
RTT connections, and its window-close mechanism adapts to the
perceived network congestion optimal point only, which is not
sufficient to meet the underlying premise of Diff-Serv architcture.
In response to these two weaknesses, this paper proposes a set of
new mechanisms for TCP's congestion control algorithm that
are specifically tailored to the Diff-Serv architecture.
While preserving TCP's ``linear increase and multiplicative
decrease" principle, these mechanisms make TCP more robust
and precise in adjusting its sending rate to network congestion,
as well as to a pre-defined service profile. Simulations and
testbed implementations are used to qualitatively demonstrate
the results. The results show that combined with Diff-Serv mechanisms
in routers, the mechanisms in endhosts can allocate resource
among TCP connections in a fair, precise and differentiated manner.
The paper also discusses incremental deployment issues and proposes
a deployment strategy for the proposed mechanisms.