Software Coordinating Committee Conference Call
		    March 14, 2002 3:00 PM EST
			  Recorder: C. DeTar

Present:

   Brower, DeTar, Osborn, Holmgren, Pochinsky, Watson, Simone, Gottlieb,
   Edwards, Mendez

Late:

   Mawhinney (other meeting)

** Action items

Agenda:

      #1. Report by Don H (or Chip W) on the formation of a new
          Jlab/FNAL working group on cluster software development.

      #2. Status report by Chip W on MP-API documentation and by Robert E on 
          C implementation and by Bob M on C++ implementation.

          Remember we are to publish this on www.lqcd.org by APRIL  1 

          Are we on track?


      #3. Level 2 API work:

          Where do we stand on Level 2 Design?
            
            How do we proceed to establish a time table for Level 2 
            implmentation(s) 
              for SZIN  (Robert E.)
              for MILC  (Carleton D.)
              for CPS   (Bob M.)    ?

           I would like to establish a reasonably detailed  schedule
           and a description of who does what when by mid April on
           implementing the QCD-API.

Minutes:

     #1) Watson: Formalize collaboration between JLab/Fermilab
         on cluster development - particularly operations software.
         DeTar: Utah people could help here.  We are running
         a largish cluster ~400+ nodes with the Maui scheduler
         and QBank accounting.
         Brower: Should give ourselves recognition for these
         collaborative efforts in the Quarterly Report
         Gottlieb: Make contact with PACI efforts.
         Brower: Also look for other SciDAC tie-ins.

     #2) Brower: We need James to test the MPI implementation
            and vet the documentation.  
            We need someone at Columbia doing the same with the
            C++ interface.
         Watson: Myrinet engineer will analyze our design
            document and gm implementation and come back with
            suggestions and even firmware, possibly, to implement
            them.
         Brower: Please send an e-mail with names.
**       Edwards: Will check the current MPI library to make sure that
            everything has the correct names and is consistent with
            documentation.
         Brower: Need a portable C++ interface.
         Edwards: Ji can do it.
         Brower: Make sure Columbia is in the loop
         [Mawhinney signs in]
         Brower: Can Columbia provide the MPI version?
         Mawhinney: Native QCDOC C++ calls are in place.
            Shouldn't the MPI version be done by someone who did
            C on MPI?  Columbia must map the C implementation on
            QCDOC.
**       Brower: So let's have Ji do it.  Columbia (Chulwoo) should
            send the C++ specs to him.  Shoot for March 28 for
            having everything ready for the group to examine.

     #3) Level 2 discussion for a subcommittee

         This portion of the call was dedicated to the Level 2
         subcommittee discussion.
         [Watson, Mendez, Gottlieb, Holmgren sign off]
         [Brower, DeTar, Osborn, Pochinsky, Edwards, Mawhinney remaining]
         Mawhinney: Proof of lazy evaluation consistency - maybe just
            a specification and a run down of all the cases that may
            be exceptions.
         Edwards: Exit level 2. Don't we need a global exit for
            communications on QCDOC,
         DeTar: But we should still have a way to remove and
            restore individual fields for efficiency.
         Pochinsky: Can this be done without treading on the hardware state 
            or state information?
         Pochinsky: Need a separate "finalize" and "extract".
         Brower: How to extract an SU(3) matrix element?
         Pochinsky: What if we want to store all the real parts
            of an SU(3) matrix element together.
         DeTar: Then we would need an accessor for every single
            real and imaginary part of every matrix element
            of an SU(3) matrix.  Sounds grotesque.  Do we really need
            to pay such a price to gain complete layout flexibility?
         Mawhinney: If we check out a field, one has to clear
            the state information - wouldn't there be a 
            performance penalty?
         Edwards, DeTar: Not any more than would happen if we tried
            to start a new shift on that field.
         Brower: Can we say we lay out data as a Cartesian
            product of data and sites?
         Edwards: We wanted to allow plenty of flexibility for
            laying out Level 2 data, so we are forced to using
            accessors.
         Osborn: What if we provide accessors that present results
            in the Cartesian product as suggested, and let
            the implementer make the efficient choice?
         Brower: Can this subgroup call in next Thursday?
         Mawhinney: Will be at Pittsburgh next week.
         Brower, Edwards: Perhaps interact in pairs.
 
      Level 3
 
         Mawhinney: Progress with Dirac ops on QCDOC ASIC Full
            multinode Dslash is running, but not global sums.  2^4
            Wilson 49% of peak w/ ASIC sends over SCU's back to itself.
            2^2x4^2 57% (Boyle); 2^4 naive staggered 36% w/o global
            sum (Christian)

Conference concluded at 5:00 PM EST.
Next general call March 28 3 PM EST.
Level 2 subcommittee call March 21 3 PM.