parAgent (auto/assist parallelizing program) anyone?

Douglas Eadline deadline at plogic.com
Mon Aug 7 16:54:08 EDT 2000


On Mon, 7 Aug 2000, Jimmy Eng wrote:

> Sorry - the correct URL is
>    http://darwin.ee.iastate.edu/~srl/
> 
> Jimmy Eng wrote:
> > 
> > Does anyone have any experience or comments/thoughts on
> > Parallelization Agent (parAgent)?  It's software to assist
> > in parallelizing fortran77 code developed at Iowa State:
> >    http://darwin.ww.iastate.edu/~srl/
> > 
> > As much as I can tell from briefly flipping through the web
> > pages, there are similarities to Bert77 from www.plogic.com
> > mentioned in the Beowulf FAQ.

I have looked at the parAgent pages and can give a quick summary
of the differences. 

1) BERT uses profile information about the target hardware.
   Profile information is used so BERT can make good decisions about
   what concurrent parts of the program SHOULD be run in parallel
   for a given hardware platform. In general, not all concurrent
   parts of a program produce speed-up when run in parallel.

   It is not clear how parAgent will decide (other than the user)
   what parts to parallelize. 

2) BERT uses a thin layer to interface to either PVM or MPI.
   BERT currently can choose from two "parallelization" 
   models (dataflow and dynamic) depending on the application
   and the profile information. BERT works on clusters.

   parAgent only supports the IBM SP and Cray T3 with an
   SPMD model.
   
3) parAgent seems to use some heuristic knowledge about
   the program (provided by the user) while BERT pretty much
   works "below the algorithm".

4) BERT can give pre-runtime estimates of parallel efficiencies.
   This can be very effective for both old codes and new ones
   (new algorithm development) with out having to run codes.

We will be introducing a static (data sliced) model for BERT
later this year.  From our experience with, old codes (sorry legacy)
the important thing is to determine where your parallelization efforts 
will pay-off.  BERT allows this question to be answered first instead of
investing several weeks of trial and error.

I hope this helps. But I think it is difficult to compare these types of 
tools unless you play with them a bit.

Doug

-------------------------------------------------------------------
Paralogic, Inc.           |     PEAK     |      Voice:+610.814.2800
130 Webster Street        |   PARALLEL   |        Fax:+610.814.5844
Bethlehem, PA 18015 USA   |  PERFORMANCE |    http://www.plogic.com
-------------------------------------------------------------------


_______________________________________________
Beowulf mailing list
Beowulf at beowulf.org
http://www.beowulf.org/mailman/listinfo/beowulf



More information about the Beowulf mailing list