Review: Congestion Control for High Bandwidth-Delay Product Networks

Paper: Congestion Control for High Bandwidth-Delay Product Networks

Name: Robert Danek.
Course: CS2209, Fall '06.

    This paper approaches the problem of designing a congestion control
from the point-of-view of being able to design the mechanism from
scratch. This differs from earlier approaches we studied that tried to
fit into the existing design of TCP.

    In particular, the paper is concerned with congestion control in
high bandwidth-delay product networks. It turns out that TCP becomes
unstable as the bandwidth-delay product increases. This is true
regardless of the queuing mechanism used for controlling congestion in
routers. In addition to lack of stability, TCP also suffers from
inefficiency on high bandwidth-delay product networks.

    The protocol that this paper presents is a replacement for TCP, and
is called the eXplicit Control Protocol (XCP). It is a window-based
congestion protocol, like TCP, but has some notable differences. Rather
than turning on a special bit of data in outgoing packets from a router,
as is done in the ECN scheme, XCP involves sending the degree of
congestion.

    The authors introduce the concept of decoupling fairness and link
utilization. Utilization refers to how much spare capacity of a link is
being used or wasted. Fairness refers to whether the different flows
going through a link are being assigned equal amounts of bandwidth. XCP
adjusts its aggressiveness based on the amount of spare capacity in the
link, thus controlling utilization. XCP also reclaims bandwidth from
flows that are taking too much, thus allowing for fairness.

    The authors run a number of simulations to show that XCP outperforms
TCP. In fact, with XCP, packets are rarely dropped. Given that the
authors simulated XCP in both conventional and high bandwidth
environments, the results of the simulations demonstrate effectively
XCP's superiority.

    Since XCP involves sending more control state information in the
packets, one could potentially criticize the protocol for adding extra
overhead to data transferred on the network. This means that the amount
of user data that can be transferred may suffer. However, given XCP is
targeted for high bandwidth-delay product networks, the network itself
will have increased capacity to handle this extra overhead, so it
shouldn't be a problem.

    Another positive feature of the paper is that the authors suggest a
scheme for gradual deployment. Even though XCP is designed as a
replacement to TCP, a mechanism is suggested for having them coexist.
This will substantially decrease the pain in migrating to XCP.
Received on Mon Oct 02 2006 - 18:24:59 EDT