Transport protocols and bulk file transfer

[stanislav shalunov <>]

                      Bulk Transport WG Meeting
                     Monday, September 19, 2005

Attendees:
        Stanislav Shalunov 
        Steven Senger
        Chester Ruszczyk
        Eric Boyd
        Cecelia Dove
        Rich Carlson
        Eric Nadel 
        Larry Dunn 
        Injong Rhee
        Susan Evett (scribe)

Meeting started at 1:30 p.m. EDT. 

Agenda
1. Review of `Summer of Code' (Google) Projects
  - Overview of projects and accomplishments
  - Next steps

2. SC2005 demo

3. Future call/meeting schedule

4. Review of previous action items:

  a. All (except Stas, Jeff, and Chet): Review the draft API Steven
     sent out (v.6.) and get comments to Steven.  [Ongoing.]

  b. Injong: track the COMNET call for papers and keep the group
     posted when a date has been published.  [Ongoing.]

  c. Injong: Run CFP for congestion control protocols past the
     transport list and send it to the general public (e.g.,
     e2e-interest).

  d. Gu: to talk to Bob about license.

  e. Injong: Pushing out the design document.

1. Summer of Code Projects related to Transport -- 3 projects that
were outstanding successes, 3 that did well, 2 that did OK and 2 that
didn't make it (one wasn't related to Transport -- Gaim plug in -- and
the other was a student who dropped out).  Outstanding success:

1. time-keeping library based on TSC register -- get time of day
without going to the kernel and switching context back and forth; this
can reduce CPU use by a significant fraction if one needs to fetch
timestamps for every packet one sends.  That hurdle is now removed --
fetching timestamps is `free' in terms of CPU cycles.  Will be useful
for any tool that needs to fetch many timestamps (delay measurement
tools, such as OWAMP & thrulay, among others), code profilers.  Good
project with wide applicability.  LD asked for where code is stashed;
SS reported that he's writing up a report on the projects as a whole.
(This one is at http://fasttime.sourceforge.net/.)  When report is
done, he'll send to list.

2. characterization of delay as packets pass through the network.  If
you want to measure network delay, you have to subtract the noise from
the observed delays.  Previously there was no characterization of
intra-machine delays; this project accomplished a lot in the time
allowed, but on FreeBSD and Linux and not on Windows.  The picture
that emerges is that they are fairly similar; no dramatic difference
between the OSes -- change is more often based on load on machine,
etc.  Difficult to say, with any certainty, what the analytical form
of the delay distribution is.  Great data collected; the sets are
available.  More research can continue from this point.

3. Work on the actual transfer tool that DIDN'T start with UDT -- if
you excluded the UDT code base, the student make more progress.  The
student had a code base that he had started several months before at
ANL that seemed to be very useful.  Student has revamped the code to
support the API, changed the architecture, and improved performance.
The congestion control piece is missing.  Student will continue to
work on this and group will see more output on this area; by end of
calendar year, should be in much better shape (this is already
usable).  (Note: student is on the mailing list but is currently on
vacation.)

Three projects on building the transport tool, two trying to extend UDT
to support the bulk transport API (per Steven S's document) to support
pluggable congestion control models.  One of the students provided a
long list of problems encountered with using UDT; based on the
problems encountered by the two students, SS feels the group has given
UDT a good try but it didn't work too well when things needed to be
changed (stand-alone, it worked very well).

Successful projects:

1. Two students working on Thrulay (a network measurement tool,
   similar to ttcp, netperf, iperf, etc., that measures delay and
   throughput) that made significant progress.

2. Project working on the time-keeping library -- Windows support
   especially important; not as robust as the Linux & FreeBSD work.

Next steps: Stas to circulate code pointers.

Injong suggested that Stas circulate the code and his report to the group.

2. SC05 Demo: Stas asked Steven if they still have a chance to run a
demo at SC05; Steven feels that there is a chance but he feels he will
have to do the primary work in getting the demo together.  Stas asked
if it would help to have Ivan look at the integration or should he
just continue to plow forward on the generic library work.  Steven
feels it would help him most if Ivan continued with the generic work.

3. Future Call/Meeting Schedule

Stas asked if the group felt it should transition to as-needed basis
for the calls and plan to meet regularly at Internet2 meetings.  Chet
felt that it would be useful to transition away from regular calls;
Larry agreed but asked that meeting times/dates/locations for
face-to-face meetings will occur.  Stas agreed to send information on
this as soon as he gets news of the meetings.

4. Review of previous action items: (see `e' above) Injong started
with an update on the paper submission status -- group needs to
continue thinking about how to publish the paper. At this point, he is
still waiting for PFLDnet information; Injong feels the group might
want to consider going out and looking for another venue(s). Stas
asked for recommendations; Injong suggested CCR and IEEE Networks.
Stas noted that the group needs to `formally declare' that it is `done
with the design document.'  He said it would make more sense if that
`announcement' came from Injong -- the `last call' period has expired
and the group needs to `announce' that the document is final.

In re `d', where Gu would talk with Bob about using UDT code but it
now appears (based on data collected in Summer of Code) that the code
base hinders progress vs. allowing progress (fine for stand-alone but
modifications are difficult).

In re `a', the API will remain a draft until implementation has
progressed further; Stas suggested that Ivan send changes to Steven as
he progresses in code.

Larry asked Injong if he ever got feedback on the NeTS submission;
Injong sent the information / comments to the group involved -- he
will send them to Larry but doesn't want to send them to the list as a
whole.

ACTION ITEMS:

a. Stas to send out a copy of the report on Summer of Code when he is
   finished.  He will also send out pointers to the code for the group
   to begin looking it over.

b. Group to review code.

c. Injong will get information on other paper publication venues to
   the group by 9/30.

d. Injong will declare (via the mailing list) that the design document
   is `final'.

The next meeting will likely be at the February Joint Techs Workshop
but exact dates and locations will be determined via email.

Meeting ended at 2:25 p.m. EDT.

-- 
Stanislav Shalunov              http://www.internet2.edu/~shalunov/

.signature: I/O Error