Software Coordinating Committee Conference Call
		     February 22, 2007 1:00 PM EST
			  Recorder: C. DeTar


Present: 
  Brower, DeTar, Levkova, Pochinski, Scholz, Joo, Edwards 
  Gottlieb, Fowler, Simone, Holmgren, Efstathiadis, Mawhinney, Basak,
  Clark, Jung, 

Absent:

   Watson, Osborn, Zhang, Khoriaty, Renner, 

=====================================================================
** Action items

Preamble

COMMENT: Please send me slides to post on
http://super.bu.edu/~brower/workshop  This website is in our
progress report to the DOE so it is good to have it as complete
as possible.


NEWS: David Richards and I were at the SciDAC kickoff meeting
      Feb 5,6 (see http://outreach.scidac.gov/kickoff/).

      All and all quite interesting. Clearly a major focus
      was to see how the  SAPs and C&I can link up and collaborate.

TRANSLATION:
    SAPs = Scientific Application Partnerships (This is what we are!)
    C&I  = Centers for Enabling Tech & SciDAC Institutes (new SciDAC2 idea)

    There are many potential opportunties in Performace (PERI), Alogrithms
    (TOPS), Graphics (Visualization & Data Management), Threading technology
    (Computer Science Institute), Training (SciDAC Outreach Center),etc

    Too many in fact but we  should exploit  these links when they
    are of mutual advantage.  In fact we are already doing this
    to a considerable degree so we should advertise this. This is certainly
    part of the new SciDAC paradigm, so it "pays" to get with the program.

  Fowler:

    As far as PERI is concerned...
    Chroma is the poster child for hard-to-analyze performance.
    There is one MS thesis based on this and possibly a paper.
    In one test 59% of time was spent in template code.
    The Livermore team is working on the analysis.  Pat Worley is the driver.
    Whether we will be able to improve performance remains to be seen.

  Brower: There is a group working on security.  They could help us
    with a Wiki page.  They might also help us with a training
    workshop.  See

       http://outreach.scidac.gov/

============= Agenda =================


1. Plan for miminal CRE implementation:
   (Results of meeting by Chip/Balint, Don, Statos)

  Balint: We discussed a common documentation environment as opposed
    to a common runtime document.
    We would standardize where it is easy to do so.
    e.g. we wanted to unify the actual run command.
    We agreed to provide generic builds of the various SciDAC packages.
  DeTar:
    The idea was to have further standardization be driven by
    the user, rather than by management.
 
2. News on Asqtad/RHMC on QCDOC?
   (Chulwoo, Carleton)

  DeTar:
    We are making progress.
  Jung: 
    We sorted out the problems with switchable precision.
**  We also needed input from Mike Clark.

2b. Discussion of a propagator format standard.

  Mawhinney:
    We would like to set a standard for propagator storage.
  Edwards:
    As for the binary format, we store all 12 spin indices of sources.
  DeTar:
    That wouldn't be desirable for our big lattices because of memory
    problems.  We would want to split the propagator into one record
    per source
  Mawhinney: 
    To specify the source, what if we created a catalog that
    describes the source and use an index: source 1,
    source 2, etc. to identify which one we use.
  DeTar:
    What if we kept the source in the file as a record, unless
    it is really trivial?  Then the metadata that goes with
    the source could describe the parameters used to build it.
  Edwards:
    Our propagator metadata has all the gauge field metadata and lots
    of other data about smearing.  We can reconstruct the propagator
    from information in the file.
  DeTar:
    We should pick a standard place in the file to put the metadata
    and allow collaborations to do what they want there.  Then we need
    a place to publish the explanation of the metadata in enough
    detail that someone else could reproduce it.
  DeTar:
    We should have a subcommittee look at this.  Perhaps Robert, Bob,
    Jim and me.
  Mawhinney: 
    It is a matter of urgency for us.
  DeTar:
    We should at least decide on the storage order and the place
    to put the metadata.  I'd hope we could do that quickly.
  Edwards
**  I'll send around the lime contents of our file and description of the
    storage order.
  
3. Shared Memory Model for multi-core
   (Don)

4. QMT Mult-Treading interface design and prototyping:
   (see QMT slides on http://super.bu.edu/~brower/workshop
        or http:/super.bu.edu/~brower/scc )

  Edwards: 
    Jie has been working on QMT, but I haven't spoken with him lately.
    Our near erm exercise was to test performance of the Wilson
    Dslash and compare MPI on separate cores vs threaded.
  Brower:
    We have to look at the far term as well where we will have 
    dozens of cores.
  Holmgren:
    On Opterons you have to make sure you use local memory
    or you take a hit.  We have been looking at simple loops
    and comparing openMP and locking processors.
    You have to touch each page of memory before starting
    on a thread.
  Edwards:
    Have you talked with Jie?
  Homgren:
**  Not yet?  I will.
  Fowler
    I'd like to volunteer UNC folks to look at performance.
    We have done benchmarking and allocation experiments and
    are aware of the problem with local memory.  As the
    number of processors increases, the ratio of local to remote
    goes down.  The hypertransport coherence protocol is not very 
    scalable.
  Gottlieb:
    Is there any literature?
  Fowler:
    No. It depends on the motherboard.
    John Miller Cromey (?) from Rice is working on multicore 
    optimization.  I'll ask him whether they could help.
  Brower:
    He mentioned SILK as a lightweight method to fork and
    spawn threads.  Andrew borrowed the thread locks (barriers)
    to do something similar with Solaris.
  Fowler:
    Perhaps making contact with the ISCADS folks would help.
**  I'd be happy to act as a liaison.

5. Et al

Committee conference concluded at 2:20 PM EST.  
Next call Mar 1 at 1:00 PM EST

======================================================================