23Oct Pooma Meeting Notes

[Jeffrey Oldham]

Participants in this week's Pooma meeting:
Jim Crotinger, Scott Haney, Mark Mitchell, Jeffrey D. Oldham, Stephen Smith

LANL decision on Pooma released Thursday.  Scott Haney is optimistic.

Proximation people will minimize Pooma work this week.

Measurements:

  Scheduling takes much more time than expected, but the measurements
  are too gross to determine the exact area of waste.
  Pooma::blockAndEvaluate() is taking a lot of time.

  Jeffrey also worries about excessive variability in execution times
  but does not know whether this is Pooma or the underlying machine.
  Scott suggested asking the Blanca people to run some tests on a
  reserved section of Bluemountain.

  Both Scott and Mark want Pooma classes to profile the code using
  Tau.  The classes should compile away when not used.  Contact John
  Hall (via Jean Marshall 505.665.4486) for the classes developed by
  Blanca.  See also the r1 profiling.  The profiling should be
  semantically based in fine enough granularity to differentiate
  different portions of code while coarse enough to avoid profiling
  skew.

  (In subsequent conversation between Mark and Jeffrey, Mark explained
  that 1) the profiling code can remain in Pooma at all times and 2)
  the profiling code can be added to and maintained by hand.  Jeffrey
  did not realize either of these were permissible.)

Optimizations:

  Ensure all fields are turned into arrays.
  STEPHEN, will send email with details.
  Clues:
  o Compressed fields on the left-hand side already have this
    implemented, but regular fields do not yet.
  o Use a zero-based field view to turn into an array.

  Deferred: Cache guard cell fills.
  Clues:
  o Look for fillGuard() or fillGuardCells().
  o Store in multipatch engines.
  o Need a mechanism to save and retrieve iterates.

Thanks,
Jeffrey D. Oldham
oldham at codesourcery.com