From: Karthik Gopalratnam (karthikg@cs.washington.edu)
Date: Wed Oct 20 2004 - 01:21:02 PDT
Congestion Control in High Bandwidth-Delay Product Networks - Katabi et. al.
This paper introduces an alternative to TCP that can provide much better
congestion control mechanisms. In addition, this protocol is aimed at
increasing efficiency when recovering from congestion control.
The paper proposes XCP, which aims to satisfy two goals:
1) Remove the instability inherent in TCP under congestion,
especially in high andwidth-delay networks which the authors' measurement
study proves exacerbates the problem, thereby suggesting that the TCP world
is destined for some form of meltdown, with high bandwidth-delay networks in
the future.
2) Provide better efficiency by solving TCP's slow restart additive
increase policy, which significantly decreases network utilization.
Another desirable consequence of the design of XCP is that the
congestion control mechanism is separate from the fairness mechanism,
thereby expanding the space of fairness policies that can be implemented in
the network.
These two goals are met by having the network provide Explicit Congestion
Notification, not in the form of just a simple boolean bit, but a window
that indicates the extent of congestion. This enables the hosts to deal with
congestion in a predictive manner, allocating bandwodth based on prediction
of RTTs in an efficient way.
The authors have done a very thorough job of evaluating the various
performance characteristics of XCP, and the results seem very encouraging.
Given that the protocol seems pretty backwards friendly, it would be useful
to actually try to deploy it in the internet.
This archive was generated by hypermail 2.1.6 : Wed Oct 20 2004 - 01:21:04 PDT