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


Present: 
  Brower, DeTar, Levkova, Fowler, Joo, Edwards, Jie Chen, Efstathiadis, 
  Scholz, Renner, Basak, Gottlieb, Osborn, Pochinski, Jung, 
  Simone, Holmgren

Absent:

  Mawhinney, Clark, Watson, Zhang, Khoriaty,  

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

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

1. Propagator format standard
   (Robert, James, Carleton et al)

  DeTar:

    I circulated a propagator format among a subcommittee of the
    affected parties.  On this call we agreed on these standards:

    File XML: 
       File layout: Use character strings to distinguish them.  For example
        USQCD_DiracFermion_12 for 12 solution vectors
        LHPC_DiracPropagator  for the dense LHPC format
        For staggered propagators, to be determined

      Gamma matrix convention
        Use character strings to specify the gamma matrix convention.
        It sounds like we all use the "DeGrand-Rossi" convention.
        Call this the USQCD convention.

      Record order for the 12 solution file:
        Pick one and require it.  Then we don't need to list the contents.

    Datatype name defines the byte order:
       Use USQCD_F3_DiracFermion
           LHPC_DiracPropagator as a synonym for the current "Lattice"
           USQCD_F3_ColorMatrix as a synonym for the current QLA_F3_ColorMatrix
   
2. QMT Mult-Treading interface design and prototyping:
   (Robert et al see QMT slides on http:/super.bu.edu/~brower/scc )

   Pochinski: We should postpone our discussion until we know more
     about the BG/Q.
   Edwards: We can make progress with our Opterons right now.
   Simone: We just need to be sure our implementations are flexible.
   Fowler: What benchmarks can we use?
   Simone: Wilson Dslash sounds like a good starting place 
   Fowler: We would like to run some experiments.
   Edwards: After the inverter, global reductions are another problem.
       The list goes on. So picking a simple test case is what we should do.
   Edwards: I circulated a document last night, following our BU
       discussion last Fall.  Jie Chen has a QMT implementation.

         [ See thread_vs_multiproc.txt ]

   Chen: We are experimenting with OpenMP for threading.
   Simone: But can you bind a process to a core?
   Chen: No.
   Fowler: In typical code there is a lot of sequential stuff between
     the multithreaded loops that become rate limiting.  I haven't
     seen impressive success stories in OpenMP.
   Chen: Threading works well up to four cores.
   Edwards: The basic question is whether it outperforms the 
     dumbest multiMPI implementation.  We don't know, yet.
   Joo: We can do timing easily, but the outcome depends on how you 
     tweak it.
   Osborn: Couldn't we run a simple test?
   Chen: I have been running only fundamental tests, but not with any
     physics application built on top of it.
   Brower: The last item in the All Hands meeting is a software 
     discussion.  This might be a good opportunity to alert
     the community about what is coming for multicore software.

        [ The call was concluded because of an audobridge conflict.
          The discussion will be continued in two weeks when
          Robert is again available. ]  

3. Shared Memory Model for multi-core
   (Don)

4. Et al


Committee conference concluded at 2:10 PM EST.  
Next call Mar 8 at 1:00 PM EST

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