Transport protocols and bulk file transfer

[Simon Leinen <>]

Stanislav and all,

I'm not sure I'll be able to make today's conference call, so I'll
introduce myself on the list

>         * Several people have joined the mailing list.

>         * Who are you and what's your interest in transport?

I work in the "network" (backbone engineering & operations) group at
SWITCH, the Swiss education & research network.  We operate a backbone
based on inexpensive GigE/10GE routers (Cisco 7600 OSR) over a simple
optical (DWDM/CWDM) transport network that we also build and operate
ourselves over dark fiber.  My specific interest in transport
protocols/applications is motivated from two directions:

1.) Our backbone routers have much smaller buffers than your typical
    Juniper- or GSR-based backbone.  The "Sizing Router Buffers" paper
    at SIGCOMM 2004 gives me confidence that we could fill these links
    quite nicely with "normal" traffic.  But I'd also like the
    occasional bulk data transfer to be able to fill the capacity of
    our links.  Ideally, those users from the "Petabyte crowd" could
    use a transport protocol that performs well for them (even on our
    "underbuffered" network) and coexists nicely with other traffic.

    (Note that this is somewhat contrary to the predominant thinking
    in the European research networking space, where those users would
    each get their own network composed of "user-empowered lightpaths"
    and could then use transports that don't have to worry about other
    traffic.)

2.) Another activity in the GN2 (GEANT successor) project is the
    "PERT", which should help users with complex (domain- and/or
    layer-spanning) performance issues.  We'll also build a "knowledge
    base" to help users solve or avoid these problems themselves.

    From our work so far, it is clear that "tune your TCP send/receive
    buffers", while helping in many cases, is not sufficient (and
    sometimes not feasible) in all bulk-transfer situations.  So it's
    useful to look at the larger issue of transport protocols,
    libraries, and applications for huge data sets.

I'm also interested in the robustness of transports with respect to
common types of packet reordering, and the ability of utilizing
multiple paths (ECMP).

Another area that I find promising is to go beyond point-to-point data
distribution.  For instance, BitTorrent seems to solve extremely well
the problem of making many copies of a large file to different places
in the network.  To me, this looks like exactly the problem that the
CERN LHC/DataGrid people are trying to solve (in a traditional way)...

We have a few Lintel machines with GE (and soon some with 10GE)
connections that we use for active measurements (IEPM, OWAMP etc.).

Regards,
-- 
Simon.