Software Coordinating Committee Conference Call
		       April 8, 2004 3:00 PM EST
			  Recorder: C. DeTar

Present:
  Brower, Gregory, DeTar, Osborn,Holmgren, Simone, Watson,
  Mawhinney, Edwards, Jung, Pochinsky

Absent:

   Mendes, 
   Neff, 

** action items
=====================================================================

QCDOC progress

   Mawhinney:

     Hooking up five motherboards.  Have 500 not completely tested
     daughter boards.

QCDOC Benchmarking

   Gregory: Has been running on xlc all week.
      Speedup is 10 - 15% over gcc.  Single and double
   Jung: Looked at the assembly file generated with new xlc
      It seems not as good as with old xlc compilation.
   Mawhinney: We are trying to understand what could be
      the problem.
   Jung: With the CG routine and improved force code, better memory
      allocation could give more than 10% on 4^4.  Temporaries
      should be in EDRAM.
   Mawhinney: We may need to call IBM about the demo Linux xlc compiler.
      We built the run time libraries with gcc so we need ELF binaries.

ILDG Workshop

   Watson:  Middleware will be light weight.
   Brower: Do we agree that Chris Maynard should speak about metadata?
**  [there was unanimous agreement]   I will send e-mail.

 

QMP till we drop!  	Clarifications,	Modification and  Extenstions
      (see http://www.jlab.org/hpc/QMP/QMP-2-changes.htm)

   h. Specialized types.

        Use simple types. QMP_XNO_ulong for the 32-bit.
        
      Strides: Can they be negative?
         Jung: QCDOC hardware uses only positive stride.
             We would have to treat it in QMP as a special case and 
             it could hurt the performance.
         Osborn: I can see a use in QDP.  Why not allow it
             so it can be used on platforms that would support it.
         Watson: Could it be implemented on QCDOC with a
             large address so it wraps.
         Jung: I don't know.
             We should have an option to warn the user of
             inefficient code.
         Mawhinney: It is difficult to see that a negative stride with
             many blocks would be a common.  One could implement rare
             cases with send multiple.  So why build it into QMP.
**       Watson: Let's table until we know whether QCDOC can wrap.
             Stride is allowed to be negative until further notice.

   i. Threads and SMP
         Osborn:
           Let's make it MPI-like.
           QMP_init_message_passing should take two parameters,
           one with the request and a pointer to take the reply
         Watson:
**         I'll ask Jie.

   j. Global reduction: require commutative and associative?
         Jung: Commutative is not required.  But associative should be.
            MPI is like this(?). Ordering is rank order if noncommutative.
**       Watson: We'll make it so.

=====================================================================
Committee conference concluded at 4:45 PM EST.  
Next plenary call April 15 at 3 PM EST.