From: Andrew Putnam (aputnam@cs.washington.edu)
Date: Mon Oct 18 2004 - 00:01:27 PDT
Congestion Control for High Bandwidth-Delay Product Networks
Dina Katabi, Mark Handley, and Charlie Rohrs
Summary: The XCP protocol is introduced as a substitute for TCP that
allows for better congestion control through dynamic congestion status
monitoring and adjustment. The paper separates efficiency control from
fairness control, making the implementation more flexible.
One key contribution of this paper is the separation of the congestion
control mechanism (also called the Efficiency Controller (EC)) from the
fairness control mechanism (FC). In TCP and many other network
protocols, these two mechanisms are inseparably blended together. XCP
separates the mechanisms, making their tasks clearer and easier to
implement. It also allows changes in one mechanism without requiring
changes in the other. For example, fairness control can be implemented
as using either even or biased fairness.
XCP improves the congestion control of TCP by using dynamic information
about the RTT of packet flows, the number of messages in the input
queues, and the available bandwidth to determine whether the link is
over or under-utilized. The congestion measurement is a variable
(rather than a boolean), so the different amounts of congestion can be
dealt with appropriately before the link becomes excessively congested.
The primary advantage that XCP seems to provide is a more rapid ramp-up
to take advantage of available bandwidth. Whether initiating a
connection or resuming a connection after it was slowed by congestion,
XCP allows clients to gain bandwidth much more quickly than TCP. This
is particularly important when dealing with short, bursty traffic that
is characteristic of much of Internet traffic. XCP achieves this by
allowing the congestion control mechanism to rapidly allocate
bandwidth, and rely on the fairness control mechanism to ensure fair
bandwidth allocation.
The authors do not do a particularly good job explaining the advantages
of XCP over Adaptive Virtual Queue (AVQ). The results do not seem to
indicate that XCP has a strong advantage over AVQ in any metric other
than bottleneck utilization, and actually does substantially worse in
average queue size and packet loss in heavy, bursty traffic. The
authors also state that XCP does not require dynamic status updates in
order to work, but there is no reason to believe that the status
updates required by AVQ are difficult to obtain, and thus there may not
be much advantage in using static values. In the introduction, the
authors claim that AVQ becomes unstable at high speeds, but they never
substantiate the claim.
The additional security of XCP touted by the authors is also
significant, though incomplete. The explicit feedback of congestion
parameters ensures that any router connected to a network node can test
the compliance of that node to the congestion control mechanism.
However, compliance with the congestion control mechanism does not
exactly equate to increased security. XCP does not seem to solve
network behavior associated with denial-of-service attacks.
The authors also do not mention how the congestion control feedback
would work for multicast or broadcast situations. It seems that XCP
would have to revert to similar congestion control mechanisms as TCP
for those situations.
This archive was generated by hypermail 2.1.6 : Mon Oct 18 2004 - 00:01:30 PDT