Internet Congestion Control for Future High Bandwidth-Delay Product Environments

Motivation: This paper proposes eXplicit Congestion Protocol (XCP) which
generalizes the ECN proposal(by informing the senders about the degree
of congestion at the bottleneck) and outperforms TCP in both
conventional and high-bandwidth environments.

Key Points:

1) XCP extends ECN by not only informing the senders about congestion
but also about the extent of congestion at the bottleneck. In XCP,
fine-grained congestion control messages are accumulated in a packet as
it travels through the network. The resulting message (feedback from
the bottle neck along the path) is passed back to the original sender in
a later acknowledgement, allowing the sender to fine-tune his rate.

2) The objective of XCP is to prevent the queue from building up to the
point at which a packet has to be dropped.

3) XCP proposes to decouple utilization and fairness control providing
more flexibility in terms of design thus enabling service
differentiation to cater to the different needs of various services. The
Efficiency Controller computes the total traffic changes over the
control interval and Fairness Controller then decides on how to divide
the feedback amongst the packets. However, the reason of employing MIMD
for Efficiency Controller and AIMD for Fairness Controller are not well
motivated in the paper.

4) The control policy for misbehaving sources is not convincing enough.
Although, the headers are more informative about the health of the
network and help in narrowing on the unresponsive flows, a separate
mechanism is still required to invoke such a check.

5) A major concern with XCP is the feasibility of its deployment. Since
the authors themselves admitted that their design rationale was not
coherent with the ideas of backward compatibility and deployment, it
dilutes the advantages of XCP. XCP can only deliver results when all the
hosts and the routers along the path support it else it would not be
able to estimate the bottleneck bandwidth effectively. Thus, XCP does
not seem to be useful for existing large scale networks.

6) The router does have to maintain a per-link estimation-control timer
that is setup to the most recent estimate of the average RTT on that
link. This increases the overhead at the routers.

7) A powerful observation is that XCP seems to be very robust to packet
loss. In most of the simulations, there are hardly any packet losses
observed.

The simulations demonstrate positive results and definite improvement
over TCP by maintaining better utilization and fairness and almost no
packet loss. However, deployment issues dilute these advantages to a
great extent.
Received on Sun Oct 01 2006 - 23:52:54 EDT