From: Shobhit Raj Mathur (shobhit@cs.washington.edu)
Date: Tue Oct 19 2004 - 21:41:50 PDT
Congestion Control for High Bandwidth-Delay Product Networks
============================================================
This paper introduces a new protocol called eXplicit Control Protocol(XCP)
which performs better than TCP with respect to congestion control. It
operates the network with almost no packet drops, and substantially
increases the efficiency in high delay-bandwidth product environments. It
also achieves fair bandwidth allocation, high utilization and small
standing queue size under varying traffic conditions.
As quoted in the paper, as the delay-bandwidth product increases, TCP
becomes oscillatory and prone to instability. This was not an issue when
TCP was designed. Today, networks with high bandwidth-delay products are
available and in future they will be commonplace. This calls for a
substitution or modification to TCP. Any modification to TCP will not
stand for long as TCP inherently has many problems which are well
described in the paper. This paper takes the right approach, by building
a new congestion control architecture from scratch while keeping in mind
the deployment and backward compatibilty issues. It also takes a more
theoretical approach using control theory to model XCP, rather than the
Ad-Hoc manner in which TCP was proposed.
The biggest contribution of the paper is a mechanism to decouple
efficiency control from fairness control. This provides a flexible
framework for integrating differential bandwidth allocations which is not
possible in the TCP framework. Allocating bandwidth to senders according
their priorities or the price they pay requires only changing the fairness
controller and does not affect the efficiency characteristics. I feel this
is what most ISPs need today and is indeed a big leap. Presently,
allocating differential bandwidth is not possible. The other contributions
are distinguishing error losses from congestion losses which makes it
useful for wireless environments. Although not flawless, it facilitates
the detection of misbehaving sources and also considers the issue of easy
deployment.
One of the immediate flaws I would like to point out, is that though the
paper wants to rebuild TCP from scratch it does not consider the issue of
security convincingly. This is a major concern in today's networks. It
proposes using policing agents located at the edges of the network to
monitor the behavior of the flows to detect network attacks and isolate
unresponsive sources. This is cannot be easily deployed and doesn't solve
all the security problems.
Overall, I consider XCP to be a viable substitution to TCP. The paper is
easily readable and substantiates its claims through detailed experiments.
It should be noted that the paper goes against the end-to-end argument by
proposing feedback from the routers to control the congestion windows of
the senders. This raises on the issue of e2e tensions as discussed in
class.
This archive was generated by hypermail 2.1.6 : Tue Oct 19 2004 - 21:41:50 PDT