Vortrag Dr. Marco Gubitoso, Univ. Sao Paulo

[Maria Cherry]

                          UNIVERSITAET WIEN 
          INSTITUT FUER SOFTWARETECHNIK UND PARALLELE SYSTEME
                            gemeinsam mit 
                                VCPC 
           EUROPEAN CENTRE FOR PARALLEL COMPUTING AT VIENNA 


      EINLADUNG ZU EINEM VORTRAG IM RAHMEN DES INSTITUTS-KOLLOQUIUMS:
                
         
       Performance models for the VOTE virtual shared memory system
 
 			   
                            Dr. Marco Gubitoso
                        Dept. of Computer Science
                 Institute of Mathematics and Statistics
                         University of Sao Paulo
                                 Brazil


                 ZEIT: Donnerstag, 2. 10. 1997, 17 Uhr c.t.
       ORT: Institut fuer Softwaretechnik und Parallele Systeme
                  1090 Wien, Liechtensteinstrasse 22, 
                         Seminarraum, Mezzanin


Abstract

Programming in shared memory systems is commonly considered to be
simpler than in message passing. However, in most cases the performance
of message passing programs is superior to the one of shared memory
programs. The main reasons for performance degradation on shared memory
systems include maintenance of consistency, barrier synchronization,
and communication redundancy.

The VOTE system provides virtual shared memory within the sequential
consistency model and allows using of less strict consistency models,
in order to increase efficiency. In the far end, it allows the use of
message passing by means of built-in functions.

We introduce an analytical performance prediction approach that assumes
unknown machine size and models some of the most crucial performance
parameters of the VOTE system, including number of page faults,
communication and barrier synchronization time.  Our method enables
detecting of performance bottlenecks and performance tuning of the VOTE
system and its applications.

Experiments with small kernels demonstrate the accuracy of our
predictions and encourages further research in the field.

If time allows, the author will also present some general results
regarding random delays in domain decomposition applications based on
asynchronous communication.