Software Coordinating Committee Conference Call
		       September 27, 2001 3:00 PM EDT
			  Recorder: C. DeTar

Present:

  DeTar, Simone, Mendez, Brower, Pochinsky, Watson, Edwards

Absent:

  Holmgren (Soudan)

Partially absent:

  Mawhinney arrived 4:02 

** = action required

1. Meeting at Jlab -
**   Brower to collect names of invitees.  Need to have advance
     list for Jlab security.

2. Progress report - 1st quarter (closes Oct 15)
**   members to send topics to be included in software presentation

3. Level 1 MP-API version 0.8 Discussion

   DeTar: What do we lose if we use the MPI scheme of identifying
          messages - i.e. message type plus sending node, rather
          than a single index in a restricted range 0 - 255?
          To set the latter up in MILC code is awkward, but can be done.
   Watson: If we want to take advantage of DMA and not involve
          the receiving host, then we have to have a reasonably
          small integer to index the receiving buffer.
   DeTar: Are messages sent differently after declareSendRelative
          and declarSendTo (DMA) if they can be sent on a grid
          to a nearest neighbor?
   Watson: No - the declareSendRelative is just a convenience for those
          who use a "virtual" grid model for the layout.
   DeTar: What happens with declareSendRelative on a switch-based
          machine?
   Watson: For switch-based machines we create a "virtual" grid
          machine - then declareSendRelative operates on
          the virtual grid.
   DeTar: We probably won't be using 
   Brower: Tasks and deadlines:
** Watson: Will send details about generateTopology
           and declareTopology
** Edwards: (1) Working on a "vertical slice": conjugate gradient 
           and would have the message passing interface
           running through MPI.
   Watson: And then (2) will provide the rest of the interface on MPI
           (3) gm implementation of the vertical slice
           and (4) general gm implementation

4. Level 1 Linear Algebra

   Edwards: What about Level 1 linear algebra?  One issue to look at:
               Operations on lists of sites - will this be useful?
               This would include gathers that might permit
               vectorization, but requires indexing arrays.
               Could operate on data in a communications
               buffer.  So would help in Level 2?

5. Return to MP API

   Mawhinney: Building a vertical slice is good for a prospective
           user who starts from scratch, but this objective
           conflicts with our other objective of demonstrating
           getting existing code running on the interface.
   Mawhinney: C++ version of Message Passing?  Not necessary
           since Columbia is starting on this.
   Watson: Has already posted something on web.
   Mawhinney: Columbia needs to have something started
           so will need to work closely with Chip (Chou Zhao sp? is
           Columbia person)
   Watson: In e-mail contact Oct 4 - 11 while traveling in Europe
   DeTar: Need to hear from Mawhinney what MILC should be 
          aiming at for the initial QCDOC implementation of the
          message passing API.

Next call, Friday Oct 12 11 AM EDT
Following call, Thurs Oct 18 3 PM EDT
Call concluded at 4:35 PM EDT.