Dr. Gilles Labonté, BSc, MSc, PhDProfessor

Department of Mathematics & Computer Science

Computations on Network Parallel Computers

Network of computers configured as a super parallel computer

Classical super-computers are very costly and few institutions can afford them.  On the other hand, these institutions (universities, hospitals, factories, schools...) all have available to them a large number of PC's and other computers, already linked together by some local network.   With the help of the appropriate software (or, for example, simply by using JAVA as we have done) these computers can be made to behave as a single super-computer, able to execute most parallel algorithms.  This network computer can comprise all computers of the network and use them in an exclusive manner (at night for example) or use only a portion of them, appropriating the relatively important idle periods of CPU that occur in normal use (instead of letting PC's run unproductive screen savers, for example).  

Artificial neural networks are data processing entities with a considerable amount of inherent parallelism.  Thus, we have endeavoured to design and test such a super-computer that would implement an artificial neural network that realize Kohonen self-organizing maps.  This neural network was then used to solve many instances of the traveling salesman problem.  

We have designed a method for distributing the loads on a network of disparate computers, on which the available processing speed can vary due to the presence of other users.  We have proved that our method is the best one possible.  Furthermore, the speedups which we have observed in our tests are in perfect agreement with our theoretical formulas.  

Our study clearly corroborate the fact that parallel computations performed on network computers are not only very economical (they cost nothing!) but also constitute a very powerful tool for research and development of applications involving artificial neural networks.

Reference: "Network Parallel Computing for SOM Neural Networks", G. Labonté and M. Quintin, to appear in the proceedings of the High Performance Computing Symposium 1999.