[GHPN-WG] Meeting notes OGF 21

[Freek Dijkstra]

In the spirit of "Anything you say will be misquoted and used against
you", here are the meeting notes.

I tried to be a bit extensive, for the people at GridNets.

Meeting notes GHPN
==================
OGF 21, wednesday, 2007 Oct 18

* Slides will shortly be uploaded to
  http://www.ogf.org/gf/event_schedule/index.php?id=959
* Attendents:
    - Jason Zurawski
    - Richard Hughes-Jones
    - Licia Florio
    - Wolfgang Ziegler
    - Cedeyn
    - Costes Kotsokalis
    - Debbie Montano
    - Ralph Niederberger
    - Inder Monga
    - Martin Swany
    - Freek Dijkstra
    - Cees de Laat

Document status
---------------
- Use cases document by Tiziana Ferrari
    - Gone through public comment
    - One favourable comment in tracker
- G-UNI draft
    - See presentation (below)
- G-OBS document
    - At external expert reviewers.
    - Richard Highes-Jones will try to get this through the
      document process.
    - Short debate on the viability of Optical Burst Switching.

Charter
-------
Cees de Laat thinks the charter needs to be updated. There is
still a need to think about beyond hybrid networking, either bottom-up
(migration from an e-mail-based "control plane" to automated lambda
set-up) or high-level thinking. By now, middleware projects start to
emerge where CPU, storage and networking are co-scheduled.

Richard Hughes-Jones: We need to make grid community aware that
the network is not a static thing that "is just there", and that
GHPN is one of few places where network people and grid people meet.

Action: This discussion is taken to the mailing list.

G-UNI presentation
------------------
Slides by George Zervas (Univ of Essex) et al., <gzerva at essex.ac.uk>
Cees the Laat presents, since none of the authors is present at the OGF.

See slides.

- Slide 8 (proposal for a generic G-UNI architecture)
  - Debbie Montano asks: who is going to implement this?
    G-Lambda, Phosporous, Enlightened need this. They probably will.
    G-UNI is an abstraction layer. A common G-UNI allows easier
    interaction between different software/ projects.
  - Inder Monga likes to standardize the communication between Grid
    users and Middleware, rather than between middleware and network
    control plane, as this slides seems to indicate.
  - A short discussion on the scope follows.
    Richard Hughes-Jones reminds use that there are standard interface
    to request CPU. Someone notes that this is not (yet) true for the
    network.
  - Richard Hughes-Jones: this is a research group, not a working group.
    It defines architecture, not protocols or implementations.
    Cees de Laat notes that the mailing list is claled GHPN-WG, instead
    of GPHN-RG for historical reasons.
  - Martin Swany: We need to reach out, to other groups.
- Slide 9: (Grid Network overlay Architecture I.)
  - The audiance and presenter are confused by the label "GUNI
    (transport)" on the lower left of the slide. It is suggested that
    the GUNI signalling is in-band, and transported over the lower
    actual network. Richard Hughes-Jones says that signalling can't be
    in-band, since signals are transmitted before a connection is set up.
    The presenter looks confused, and resolves the issue by skipping to
    the next slide.

Cees de Laat is not 100% of the status. He thinks it was already ready
for internal review by the working group since the previous OGF, but
suspects it has not gone to that status since everyone is busy.

Richard Hughes-Jones appreciates the overview that the work gives, but
sees a risk that it is overly complex. Cees de Laat agrees: It states
a lot together. Cees thinks that publication as informational document
is good basis to move forward for grid middleware projects. It shows
to grid people how to deal with network resources.

Martin Swany suggests to change the name of the document. It reads as
if it specified *THE* G-UNI, while it specifies *A* UNI. Suggestions
are: "Discussion of G-UNIs" or "A UNI study". Richard Hughes-Jones
and Inder Monga agree.

Phoebus presentation by Martin Swany
------------------------------------
- There is a gap between backbone bandwidth and the actual perceived
  TCP throughput: the "bandwidth gap"
- Single TCP streams are important (despite GridFTP, alternative
  transport protocols, etc.).
- Phoebos segments the transport. It uses OSI session layer.
- Buffering of data in the network.
- 3 segments: access network+ core + access network. TCP
  termination close to user gives better performance.
    - Martin sees a simularity with "burst switching"
- Connection negotiation, as required in this concept, removes
  need for a firewall, if authenticated.
- Martin argues that the end-to-end arguments may no longer apply
  - bring back state in network
  - Cees de Laat cares about the feedback: he wants to know for
    sure that the data is arrived at the destination, and wonders
    who closes the connections: the first hop or the final
    destination. If the first hop, how to make sure all data
    arrived at the destination?
- Martin stresses that "an adaptation layer atop the existing IP
  network is a viable path to innovation", and plugs this approach
  as a "Do-it-yourself-GENI"

Cees closes the sessions and hopes to see everyone at the next meeting.