Review: CM

Review: An Integrated Congestion Management Architecture for Internet
Hosts (CM)

Reviewer: Di Niu

This is an original and solid paper, which presents a novel framework
called CM for managing network congestion from an end-to-end
perspective. The underlying motivation behind the new scheme is that
existing TCP congestion controls are unresponsive in face of
congestion. However, responsiveness is more and more important in a
modern network for reasons such as the prevalence of multiple
concurrent, short TCP connections and UDP based streaming. These
characteristics of the modern internet demands highly on the
responding speed of congestion control protocols. By going back to
the end-to-end argument, this paper tries to circumvent the hairiness
of coordinating competing flows in the routers, as well as to promote
internet's responsiveness to congestion.

CM implements all of its congestion control at the sender and the
receiver. The basic idea is to adjust traffic at the sender based on
the feedback from the receivers. This information is contained in a
CM header. The detailed mechanisms of CM are intricate. It
incorporates a window-based AIMD scheme with traffic shaping, a loss-
resilient protocol to periodically elicit feedback from receivers, an
exponential aging mechanism to regulate transmissions in a stable
manner when feedback is infrequent and a scheduler to apportion
bandwidth to flows. The paper also presents the CM API , an API for
applications and transport protocols to learn about and adapt to
network congestion and varying bandwidth.

However, the paper is not without drawbacks. It seems when the
congestion control studies have become more and more hairy, all
trying to refine mechanisms at routers to coordinate competing flows,
this paper proposes to go back to E2E arguments. It is somehow the
dramatical contrast between CM and its contemporary protocols that
has distinguishes this paper. If it were published just after
Jacobson's original paper on congestion control, I doubt that the
value of CM may be get discounted. Actually, the paper just compared
CM with newReno which is not suitable for new scenarios. This is
certainly not enough. I think it's better to compare CM with all the
protocols that dated all the way back to Jacobson.
Received on Tue Oct 10 2006 - 00:27:19 EDT