Congestion Control for High Bandwidth-Delay Product Networks
Dina Katabi, Mark Handley, and Charlie Rohrs
Summary:
This paper presents a new internet protocol named eXplicit Control
Protocol (XCP). XCP is designed to improve congestion control in high
bandwidth-delay product networks. Such networks are expected to
increase in the future, and TCP has been shown to become inefficient
and prone to instability as the bandwidth-delay product increases. In
contrast, it is shown in the paper that XCP remains efficient, fair,
scalable, stable, and has a near-zero packet drop rate as this product
increases.
XCP uses a congestion header (adding a few additional bytes to the
packet header) to convey control information to the sender and
routers. While TCP uses packet drops to determine flow rates, XCP
uses more specific information provided by the congestion header
and adjusts flows before packets start to drop.
A key concept of XCP is its decoupling of congestion control from
fairness control. This decoupling allows more efficiency as well as
flexibility in bandwidth allocation schemes.
The authors provide pseudo code for implementation of XCP and describe
how this protocol could be gradually deployed.
Critique:
There are a number of concerns regarding the simulations that were
reported in the paper:
- while the queue sizes were always set to equal the bandwidth-delay
product, it appears that the corresponding parameters may not have
been scaled up appropriately for the competing algorithms.
- the number of flows used in most of the tests was relatively small,
and it is shown in Figure 6 that while the competitive algorithms
have poorer bottleneck utilization for smaller numbers of flows,
they all have high utilization rates for larger numbers of flows.
- there are a number of missing data points in Figure 4. These
omissions appear to be unjustified as the data exhibits some erratic
behavior otherwise.
- the authors failed to make comparisons to TCP Vegas.
The other major concern is that a critical point in the paper is that
TCP becomes "unstable" with large bandwidth-delay products while XCP
remains stable. However, the authors failed to define stability and
also failed to demonstrate that TCP oscillations are inherently
problematic.
Finally, there is a concern about XCP's ability to ensure fairness.
It appears that a sender could lie about their estimated round-trip
time and obtain an unfair allocation of bandwidth.
These points were made by CS679 participants and summarized by
Sharlee Climer