Transport protocols and bulk file transfer

[Larry Dunn <>]

Colleagues,
On our last call (~2 months ago),
I promised to write up the stray thoughts I had 
re: seeing if "the community" thought there was
value in bringing togther:
a. folks that write TCP stacks in workstations, and
b. folks taht do queuing algorithms in routers (per scott: +middleboxes), etc.

So, here's a draft of those thoughts.
We may/may_not choose to discuss tomorrow...

 *  The TCP congestion control and performance characterization
    community consists of at least two sets of researchers and
    implementers - those that develop queuing strategies for routers,
    and those that develop congestion control methods for end-systems.
    There are not many venues where these practitioners interact on a
    deep level, yet their end-products may interact significantly,
    since they are both part of the end-to-end path. There are other
    community members that have a keen interest in, or impact on
    end-to-end performance, e.g., network operators and middle-box
    designers.

 *  We propose to survey the stakeholder communities, to see if there
    is interest in convening a workshop targeted at identifying and
    understanding interactions between in-network queuing strategies,
    and end-system congestion control (CC) behavior. It may be that
    this topic has been probed, and/or that there is not sufficient
    interest in converging these groups. But it is also quite possible
    that the two communities have not had sufficient interaction in the
    past, and the larger TCP community may benefit from a broader
    understanding of the interplay between in-network queuing
    strategies and end-system congestion control algorithms.

 *  Relevant questions whose probing may benefit both communities
    include: "What are commonly deployed in-router queue management
    strategies, and what is the innovation rate in that space?"; "What
    is the spectrum of congestion control strategies (e.g.,
    sender-side-only, need-help-from-router, loss-based, delay-based,
    combinations)?"; "How much do we understand about interactions or
    dependencies between queuing and CC?"; "How do choices made by
    either community impact incremental deployment possibilities?";
    "What are tradeoffs in incorporating some level of responsiveness
    to low-rate probes in routers (vs. per-packet processing)?"; "Are
    there characteristics of some CC algorithms that make them robust
    to a wide variety of queue management strategies?"

 *  For example, one common mis-fire between (congestion control)
    researchers and  (router queue) implementers occurs when algorithm
    developers design innovative methods that require "a little
    per-packet help from the router." When this conversation is brought
    to router designers, the pushback is sometimes of the form, "Well,
    that would have to go into hardware; we could do it, but this
    better be *the* winning algorithm for the Internet, because we can
    only do it once." Or, "forget it."  Conversely, if router designers
    saw value in incorporating low-level tools that provided
    flexibility in some minimal number of high-speed per-packet
    operations, interactions between queuing and congestion control
    might flourish.

 *  Possible sponsors for such a workshop include: NSF, DOE, IETF,
    ACM-Sigcomm, and Internet2-Joint_Techs.

 *  TBD- committee, paper/participant invitation/selection,
    location, duration

<end>


Best regards,
Larry
--