From: Tom Christiansen (tomchr@ee.washington.edu)
Date: Tue Oct 19 2004 - 17:20:44 PDT
In this paper a replacement protocol for TCP is described. This protocol,
XCP provides a much more efficient congestion control mechanism than TCP,
thus, increasing throughput. In addition, the authors point out that the
slow-start congestion control mechanism in TCP is not only slow, it is also
unstable as all traffic is slowly ramped up to the point where network
congestion occurs but abruptly reduced when congestion (or more accurately:
packet loss) is detected. The implemented XCP protocol may coexist on the
same network as TCP and its fairness controller is shown to be "TCP
friendly". Unlike other congestion control mechanisms which tend to be
updates/patches to TCP, XCP is a complete redesign of the TCP protocol.
The paper is generally well-written and does provide enough scientific
background to make a compelling argument for using the XCP protocol.
However, there appears to be a fundamental flaw in the stability analysis
in section 4 and appendix B: Throughout the analysis a continuous time
model is used. As the system feedback is transmitted as part of the data
and ack packets, the system is clearly a discrete time system. While it
isn't uncommon for discrete time systems to be modeled using continuous
time models, I'm surprised that the authors didn't bother to either model
the system using a discrete time model or at least use one of the many
control systems simulation programs available to verify that their model is
valid. Especially, since they do make quite a point out of proving that
their control system is stable. That being said, their test results show no
signs of instability - something which definitely cannot be said for the
TCP examples shown.
While XCP can coexist with TCP on the same network, deployment of XCP will
require changes in routers, gateways, etc. Thus, it won't happen overnight.
For smaller under-utilized LAN's the protocol probably won't make much
difference. But for fast connections with high delay bandwidth product this
protocol should provide significant increase in bandwidth utilization.
Hence their "light that fiber" campaign...
Positive: Relatively easy to read and get a fundamental understanding of XCP.
Negative: Most of the equations break up the flow and could have been put
in an appendix.
This archive was generated by hypermail 2.1.6 : Tue Oct 19 2004 - 17:20:51 PDT