Software Coordinating Committee Conference Call
		       October 18, 2001 3:00 PM EDT
			  Recorder: C. DeTar

Present:

  DeTar, Holmgren, Simone, Brower, Pochinski, Edwards, Watson

Absent:

  Mendez
  Mawhinney

** = action required

#1  Writing the Software Section  of  Quarterly Progress Report: 
    Draft due Oct 31 --- Halloween.

    1-2 pages SciDAC report on 1st quarter (July 15 to Oct 15, 2001) from each:
**  By next week - send to all the committee
    
         Jlab/MIT              --- Chip/Robert
         BNL/Columbia          --- Bob M.
         MILC                  --- Carleton, Doug
         FNAL                  --- Don
         Illinois              --- Celso

    Chip: Report on all activities and say what fraction was supported
          by SciDAC and what fraction by other.

#2. Organizing the min-reveiw part of the Nov 8&9 Workshop
**  (You can put in your suggestion as a Postscript to the Report)

    Rich: Discussion of vertical slice.
          Obvious areas where things need to be threshed out
          List of instructions, naming conventions,
          Edwards, Pochinski, Brower: Document of what instruction
          set looks like.  Want a green light to go forward
   	  with the design.
          Also Celso's work - brainstorm.
    Chip: End of workshop - put final stamp of approval on the
          MP-API version, so it is more static.
          Also, need a round-table discussion of implementations
          MILC, CPS.
          Discussion of error handling.
          Need a discussion of our 3-year plan:
    Rich: e.g. binary file formats, analysis tools, as discussed
          in the proposal.
    Chip: And run-time environment

#3. Linear Algebra API level 1 and 2 documentation etc.
    (Rich/Robert/Andrew work on this.)

    DeTar: What if we used a list of pointers and not indices
      in the gather - scatter linear algebra primitives?
    Chip:  Might have 8 byte addresses - so could cost more -
      tradeoff with CPU address computation in the case of indices
    DeTar: What if we also included a mask?  Could be useful
      at Level 2 for distinguishing between on-node and off-node
      data.

    Edwards: Message passing - propose a distinction and 4 routines
      send to always uses FIFO a la MPI1
      send relative also always FIFO
      put to (DMA) restricted number of buffers
      put relative (DMA) also restricted
    DeTar: Are the logical machine and send (put) relative parts of
      the MP-API really essential?  To me it looks like they could be
      separated and provided as an add-on for users who want that
      abstraction.  The MILC abstraction is different, but we aren't
      proposing putting it in the MP-API as well.  It would be good to
      find out what the community wants.
    Edwards: Agree could do Level 2 without a virtual grid. (***)
    Edwards: Easy enough to include it, so why not do it?
    DeTar: Would be good to keep the MP API as lean as possible.
      Let's think about it before committing one way or the other.
      
Next call: Thurs Oct 25 3 PM EDT (Chip excused - also Nov 1)
Call concluded at 4:45 PM EDT.




(***) Robert's correction --- "not exactly": 

That is not what I said - you have to have a virtual grid somewhere
underneath level 2; otherwise, it makes no sense to communicate along
+x in a multiply(src1,shift(src2,+x),target) . Under level 2 something
has to know about subgrid sizes, etc. By the design of the API-s, it
may not be possible to implement even Message Passing without the
specific grid stuff at the level 1 MP layer - namely on a grid based
machine like the QCDOC may require grid setup. However, the impact of
the grid stuff on the MP layer is really quite small on even an MPI
implementation, and the generic user will need the functionality
anyway, so why not include it.