[Acpc-l] Vortrag: Prof. Tor Sorevik, 7.6.2001

[AURORA Project]

	   UNIVERSITAET WIEN INSTITUT FUER SOFTWAREWISSENSCHAFT
			      gemeinsam mit
	    FWF-Projekt Spezialforschungsbereich F011 "AURORA"


       EINLADUNG ZU EINEM VORTRAG IM RAHMEN DES AURORA-KOLLOQUIUMS

	      ZEIT:  Donnerstag, 7. 6. 2001, 17.00 Uhr c.t.
		ORT:  Institut fuer Softwarewissenschaft,
		      1090 Wien, Liechtensteinstrasse 22, 
	                      Seminarraum, Mezzanin
	                      
			    Prof. Tor SOREVIK
			  University of Bergen,
			   Dept. of Informatics
				  NORWAY


		Performance Issues in Parallel Programming
			       Using OpenMP


In this talk we will report on some of our experiences  with OpenMP on some
of  the  applications  we  have  been  working  on at  Parallab/  Dept.  of
Informatics  at the  University of Bergen.  Our claim is:  Parallelizing  a
sequential  code using OpenMP is easy.  Obtaining  high  performance  might
however,  be  quite  difficult.  In  this  talk  we  discuss  some  of  the
performance  issues such as load balancing,  granularity and data locality.
We illustrate  the problems with examples from real  applications  and will
discuss ways of dealing with the problems.  The standard  approach has been
to  insert  OpenMP  directives  into the code and let the  compiler  do the
parallelization.  We  will  also  discuss   some  issues   related  to  the
performance  of  OpenMP.  The first is on the  potential  conflict  between
optimal sequential performance and parallelization.  The second is the need
for  controlling  of  data  distribution  on  cc-NUMA  systems  when  doing
SMP-programming.  We will argue that data distribution  directives  similar
to those known from HPF is needed for this purpose.