Vortrag: Prof. B. Szymanski, 10.6.1999

[Maria Cherry]

                             UNIVERSITAET WIEN 
              INSTITUT FUER SOFTWARETECHNIK UND PARALLELE SYSTEME
                                gemeinsam mit 
                                    VCPC 
               EUROPEAN CENTRE FOR PARALLEL COMPUTING AT VIENNA 
                       
              FWF-Projekt Spezialforschungsbereich F011 "AURORA"


        EINLADUNG ZU EINEM VORTRAG IM RAHMEN DES AURORA-KOLLOQUIUMS
                         
            
                                               
                    Performance Analysis Tools for Parallel 
                    Object-Oriented Scientific Computations
  
			  
		              Boleslaw K. Szymanski
                         Department of Computer Science
               Rensselaer Polytechnic Institute, Troy, NY, USA 

		   
                  
                  ZEIT: Donnerstag, 10. 6. 1999, 17.15 Uhr s.t.
          ORT: Institut fuer Softwaretechnik und Parallele Systeme
                   1090 Wien, Liechtensteinstrasse 22, 
                          Seminarraum, Mezzanin



Abstract:

Object oriented technology has significantly changed the way programs
are developed and the corresponding change is needed in performance
analysis of the resulting codes. Traditional performance analysis
focuses on the control flow graph of a program. The sequential and
parallel codes differ just in the number of simultaneous paths being
traversed. Such control flow graph oriented view is insufficient for
object oriented codes.  Object oriented performance analysis inherits
also universal concerns, such as data scalability, instrumentation
intrusion, attribute derivation.  Our approach to addressing such
concerns is to map a program structure onto a database of performance
results. In addition to ensuring data scalability, this approach
enables comparative analysis across multiple parallel systems while
ensuring minimal perturbation and data scalability.  Scalability is
achieved by collecting a fixed amount of statistical information for a
subset of performance critical events in the global control flow
graph.

This approach has several advantages. Centralized repository of
instrumentation data is coupled with information that facilitates
comparative analysis.  Standard interfaces to instrumentation data
allow users to customize analysis by deriving attributes and building
specialized visualization tools.  Database schema can be augmented to
support inter-object communication among member function control flow
graphs for parallel object oriented codes.  We demonstrate use of this
approach in analyzing performance of parallel FEM codes and parallel
adaptive PDE solvers.