Rethinking Startup in Congestion Control

         Dan Liu*, Mark Allman+, Shudong Jin*, Limin Wang#
                 * Case Western Reserve University
             + International Computer Science Institute
                        # Bell Laboratories

Traditionally, TCP congestion control starts transmission by probing
for an appropriate sending rate using the slow start algorithm
[Jac88].  The notion behind this technique is to essentially begin
transmission with a very conservative sending rate that is assumed
to be appropriate for the vast majority of network paths (a handful
of segments per round-trip time).  From this modest starting point,
slow start increases the number of segments transmitted every
round-trip time (RTT) exponentially.  However, in high-bandwidth
and/or long-delay networks slow start is a time consuming process
that requires a large amount of data to arrive at an appropriate
sending rate that well utilizes the available capacity.

A number of alternatives to slow start have been proposed in the
literature.  These mechanisms fall into one of three broad
categories.  The first class of schemes attempts to use bandwidth
estimation techniques to assess the available capacity of the
network path without consuming the capacity (e.g.,
[PRAKS02]). Another class of mechanisms shares information between
connections such that subsequent connections do not have to probe
the network path for an appropriate sending rate, but can use the
sending rate learned from a previous connection (e.g., in
[PK98,BRS99]).  A final class of schemes involves all network
elements (routers and hosts) along a network path in arriving at an
explicit sending rate that the network can support (e.g., [JFAS05]).
These schemes all have their pros and cons.  In our work, we take a
different approach to the problem.  We introduce "Jump Start" in
which we make the dramatic assumption that a "start up phase" is not
required and that a host can begin transmission at an arbitrary
rate, with the major caveat that the remainder of TCP's congestion
control algorithms remain in place.  In some sense, Jump Start and
Quick Start are at opposite ends of a spectrum.  On the one end,
Jump Start is easy and makes an assumption that the network can
handle traffic without carefully probing for the exact right sending
rate.  On the other end of the spectrum, Quick Start very carefully
checks with all elements in the network to gain explicit permission
to begin transmission with a large sending rate.

Obviously, Jump Start takes the risk of using a ridiculously large
initial sending rate which may cause problems for the connection
itself and for other connections attempting to share the given
network path (or portions thereof).  We intend to present very early
results that illustrate the bounds on the positive and negative
implications of Jump Start.  Our goal is to then step back from
these bounds and attempt to identify ways that Jump Start could be
made more amenable to real networks.  For instance, what if Jump
Start simply uses the information at hand (e.g., the last mile
bandwidth) to more accurately set an upper bound on the sending
rate?  Or, what is the impact of a realistic traffic mix where most
of the transfers are short and therefore cannot attain a high
transmission rate at all?  Or, what if all segments transmitted
during Jump Start are so marked such that routers could treat these
segments differently when encountering congestion?  We are exploring
all of these questions, as well as the impact of Jump Start on
competing traffic, fairness and security.

References

  [BRS99] Hari Balakrishnan, Hariharan Rahul, Srinivasan Seshan.  An
    Integrated Congestion Management Architecture for Internet
    Hosts, ACM SIGCOMM, September 1999.

  [Jac88] Van Jacobson.  Congestion Avoidance and Control, ACM
    SIGCOMM, August 1988.

  [JFAS05] Amit Jain, Sally Floyd, Mark Allman, Pasi Sarolahti.
    Quick-Start for TCP and IP, October 2005. Internet-Draft 
   draft-ietf-tsvwg-quickstart-01.txt (work in progress).

  [PK98] Venkata Padmanabhan, Randy Katz.  TCP Fast Start: A
    Technique For Speeding Up Web Transfers, Proceedings of IEEE
    Globecom, November 1998.

  [PRAKS02] Craig Partridge, Dennis Rockwell, Mark Allman, Rajesh
    Krishnan, James P.G. Sterbenz. A Swifter Start for
    TCP.  Technical Report No. 8339, BBN Technologies, March 2002.