From: Chuck Reeves (creeves@windows.microsoft.com)
Date: Wed Oct 20 2004 - 08:12:03 PDT
The paper, "Review of Congestion Control for High Bandwidth-Delay
Product Networks" was written by a number of researchers from MIT and
Berkley. It was submitted to SIGCOMM in August of 2002. It presents a
new congestion control protocol for TCP networks. The protocol called
Explicit Control Protocol (XCP) proposes to increase the level of
coordination between sending logic on endnodes and routers by adding
additional control information in each TCP packet. The authors then
attempt to show through testing and measurement that this increased
level of coordination can result in more efficient changes in sender
behavior in turn resulting in improvements in bandwidth utilization and
fewer dropped packet due to congestion all while enabling a fair use of
the network.
In an XCP enabled channel the protocol requires each send to add an
additional packet header (called a congestion header) containing its
current congestion window and round-trip-time. The router will
periodically calculate adjustments to the congestion window and update a
packet with this feedback as it flows towards its receiving endnode. The
receiving endnode will attach this header into the acknowledgement
allowing the sender to modify its congestion window appropriately.
The main contribution of the paper is the analysis and design of the
feedback control. The decomposition of the calculation of the feedback
into 2 components seemed quite logical. As efficiency calculations were
performed using aggregated information and the feed into the fairness
calculation to determine modifications to individual packets.
The one section that I thought was particularly weak in the paper was
the description provided on how this technology would be introduced into
the Internet. I specifically didn't think they were being realistic when
the authors say "These checks can be done using simple TCP and IP
options", to determine if the existing network path and endnode supports
XCP. More practical work is needed to demonstrate how this protocol
would be deployed.
That said I did appreciate a number of technical aspects of this paper:
1. The thorough testing the authors did to contrast the behavior of
this protocol with other existing feedback schemes was the best we have
read so far. Specifically figures 4 and 5 establish clear contrasts
between XCP and RED CSFQ ect...
2. Having studied Linear Feedback Systems as a EE undergraduate the
analysis provided in appendix B and specifically the nyquist and bode
plots in figure 14 provide a formal methodology for understanding the
design approach and it's resulting behavior.
This archive was generated by hypermail 2.1.6 : Wed Oct 20 2004 - 08:12:09 PDT