An Integrated Congestion Management Architecture for Internet Hosts

Motivation: This paper presents a framework for managing network
congestion from an end-to-end perspective.

Key Points:

1) The Congestion Manager(CM) puts the application in control of
deciding what to transmit at each point of time as well as the
distribution of bandwidth amongst its flows.

2) CM should be able to support application heterogeneity as well as
traffic heterogeneity. In essence, it should let the application decide
its parameters and thus provide a framework for applications to be able
to use it according to their needs and requirements. A feedback
mechanism is also provided in order to be able to provide feedback to
the CM.

3) The CM acts as a mechanism to ensure stable network behaviour by
sharing congestion information between them. The hybrid scheme of window
and rate based avoids overwhelming bottleneck routers in bursty
transmissions and also allows for an effective feedback mechanism.

4) The sender's CM uses packet loss and ECB bit set by routers and
echoed by the receivers. A weakness seems on the choice of feedback
frequency which was arbitrarily determined to be one-half RTT. The
application can notify the sender application for providing feedback or
alternately, the sender could probe the receiver periodically for
responses. By maintaining state information about probes as well as
number of packets sent between the probes, the per-flow loss rate can be
estimated by the sender.

5) As has been debated in the paper, the CM packet header does introduce
overhead as well deployment issues. However, it has been argued that it
is essential to be able to do loss and congestion detection as well as
handle reordering issues.

6) Exponential aging provides a way to handle unresponsiveness of
probing. In each RTT after a reponse is not received within a RTT of a
probe, the open-loop transmission rate is halved. The "half-life" is
chosen by keeping track of the minimum of all round trip samples.

7) A considerable amount of statistics need to be maintained at both the
receiver as well as the sender. CM basically provides the applications
an interface which they can use to be able to utilise the available
bandwidth in a manner they prefer. However, the paper has not talked
sufficienty about fairness issues in scheduling.

8) CM framework provides a good platform for experimenting with and
deplying new congestion control algorithms.

Thus, CM enables proper and stable congestion behavior and provides a
simple API to enable application to learn and adapt to network state. It
does improve consistency and predictability of network transfers and
claims to provide benefit even when deployed at senders alone. An
interesting implication seems that one could program its usage of
bandwidth although the fairness issue needs to be dealt with critically
to make such an approach useful. Also, a considerable overhead is now
shifted to the application programmer to use such an API.
Received on Mon Oct 09 2006 - 20:20:52 EDT