Abstract

Explicit Congestion Notification and the TCP Paradigm.

Teun Ott
New Jersey Institute of Technology
CS

In TCP, packet loss is used by the source of a TCP flow to gauge
congestion and to adapt the congestion window to the level of congestion
in the network. In RED, this mechanism is used by routers to signal,
by preemptively dropping packets, congestion to the sources.

ECN (Explicit Congestion Notification, see RFC 3168, Floyd and
Ramakrishnan, Sept 2001) replaces ``dropping'' with ``marking'' as the
signalling mechanism. The most obvious advantage is that marked packets
do not need to be re-transmitted. Derived advantages are that the marking
rate can be arbitralily large, and thus that the amount of information
``signalled'' can be much greater.

This allows for jointly designing new router- and source behaviors.

The most likely payoff of such a joint design is to make it possible for
`best effort' type traffic to efficiently utilize the bandwidth left
unused by higher priority (say VoIP, Video) traffic, even if the
bandwidth available to the best effort traffic varies considerably.
(Either because the higher priority traffic varies, or because the
total bandwidth varies, e.g. as in a fading microwave channel.)

A mathematical model that illustrates the kind of investigation
required will be presented, and simulation results will be presented
that show that even with a fairly naive approach it is possible to
improve QoS to low priority traffic, without doing damage to higher
priority traffic.

Preparatory reading can be found in

http://web.njit.edu/~ott/Papers/index.html,
Square Root Law, ECN, and Milcom2001.

References:

T.J Ott, J.H.B. Kemperman and M. Mathis

The stationary behaviour of ideal TCP congestion avoidance


T.J. Ott ECN
protocols and the TCP paradigm

A. Misra and T.J. Ott
Jointly coordinating ECN and TCP for rapid adaptation to varying bandwidth