%T Performance modelling of a parallel meural network simulator
%A  Tom Tollenaere,  Dirk Roose
%E  Alastair R. Allen
%B Proceedings of WoTUG\-15: Transputer Systems \- ongoing Research
%X A model program structure is presented for parallel
   applications with local interactions between the data
   elements, such as neural networks simulations and the
   solution of partial differential equations. The performance
   of this model program is analyzed both theoretically by
   means of classical performance models, and experimentally
   using a parallel neural network simulator program. The
   program runs on a Meiko transputer array, and uses the Meiko
   CSTools libraries for its communications. The comparison of
   both analyses allows to predict applications performance on
   new and other machines, and indicate what parts of an
   application are worth optimizing. Moreover, it is shown that
   classical theoretical models not always capture the behavior
   of a real machine.