Summary: XCP

Summary: XCP

TCP becomes more oscillatory and prone to instability as the
delay-bandwidth product increases. This motivated the authors to come up
with a new congestion control protocol (XCP) that works well in high
bandwidth delay product networks as compared to TCP.(esp in high capacity
links or in Satellite links or wireless LANs )

The paper explains three important design rationale behind designing XCP.
First, authors assume loss is a poor signal of congestion and hence tight
congestion control requires explicit and precise feedback. Second, the
aggressiveness of the sources must be adjusted dynamically according to
the delay in the feedback-loop. Third objective is to make dynamics of the
aggregate traffic independent of number of flows thus leading to
decoupling Efficiency Control(EC) or Utilization Control and Fairness
Control(FC).

The authors propose XCP protocol which tries to achieve the above
objectives. Routers use packet Congestion Header, that contains H_cwnd,
H_rtt and H_feedback, to handle efficiency and fairness control. Sender
initializes H_cwnd, H_rtt based on his values and populates H_feedback
based on his demand. Router along the path modify the H_feedback(+ve or
-ve) to control congestion windows of the sources. Congestion control
decision are made at the Routers and Routers take a single control
decision every average RTT.

An XCP router uses RED queue equipped with EC and FC. The purpose of
efficiency control is to maximize the link utilization while minimizing
the drop rate and persistent queues. The important observation here is
that protocol uses persistent queue size as opposed to transient queue
size to allow for bursts. While EC computers a desired increase or
decrease in the aggregate traffic, job of FC is to apportion the aggregate
feedback into individual packets to achieve fairness. By decoupling EC and
FC, we are able to ensure that protocol is utilizing the maximum available
bandwidth while providing fair treatment for each individual flow.

With XCP, dynamics of aggregate traffic stay independent of number of
flows traversing in the link. Furthermore, since the feedback is computed
over an avg RTT, XCP becomes less aggressive as the round trip delay
increases. The paper presents extensive simulations to understand
Bottleneck utilization, Average bottleneck Queue, and Bottleneck drops
over various topologies. The results show that new protocol dampens
oscillations and smoothly converges to high utilization, small queue and
fair bandwidth allocation. The simulations to study convergence dynamics,
robustness to handle busty traffic suggests that XCP preforms
significantly better high bandwidth delay environments.

The paper presents of idea of decoupling fairness control and utilization
control that opens up new avenues for service differentiation for e.g.
paper describes a simple shadow price model for bandwidth differentiation.
Finally, it provides insight as to how XCP based CSFQ and TCP-friendly XCP
can be implemented over the Internet.
Received on Tue Oct 03 2006 - 05:47:48 EDT