From: Jonas Lindberg (jonaslin@kth.se)
Date: Tue Oct 19 2004 - 17:44:40 PDT
Congestion Control for High Bandwidth-Delay Product Networks
By D. Katabi, M. Handley and Charlie Rohrs
Reviewed by Jonas Lindberg:
In this paper, which was published in 2002, the authors suggest a new
window-based congestion control protocol: the eXplicit Congestion Protocol
(XCP). The reason for engineering a new protocol is that TCP today is
becoming less and less efficient as the network capacities is increasing.
The design of the XCP protocol was done from scratch without any concerns
about backwards compatibility or deployment, and the result is very
interesting.
The most important design differences when comparing XCP with TCP are:
* Use of explicit and precise feedback from routers to end-hosts.
* Efficiency control and fairness control is separated.
The most important results from comparing XCP with other solutions using TCP
and different router congestion control algorithms are:
* Using XCP, very few packets are dropped (difference of order 10^3).
* Oscillations in transfer rate decrease quicker for XCP than for TCP.
* XCP is utilizing bandwidth better than TCP (higher efficiency).
* Bottleneck router queues are often shorter in XCP than in TCP, which
results in less latency.
The paper is generally well written; the design choices are motivated and
the authors show awareness of problem areas, such as deployment. Results
from simulations in which XCP was compared with other solutions (based on
TCP and different router congestion control algorithms) are clearly
presented and discussed.
One thing that I miss in this paper is a motivation for the choice of buffer
size and data packet size in the simulation setup chapter. Another thing is
that I would have liked is more discussion on how to deploy the XCP. The
authors do bring this up and dedicate nearly a page to discuss it, but since
it could become a very important issue if we were to deploy it, I would not
mind reading some more on this. But may be that is another paper.
I very much enjoyed reading this paper. Discussing alternatives to TCP and
what possibilities these provide is important and helps pushing the
development. I find this paper very relevant and interesting.
This archive was generated by hypermail 2.1.6 : Wed Oct 20 2004 - 01:45:05 PDT