Trade-oos between Shared Virtual Memory and Message-passing on an Ipsc/2 Hypercube M Emoire Virtuelle Partag Ee Ou Echange De Messages Sur Un Hypercube Ipsc/2 ?

This paper presents the results of an experiment which evaluates the performance of shared virtual memory running on a distributed memory parallel architecture (iPSC/2 hypercube). Two parallel versions of the modiied Gram-Schmidt algorithm are presented. Each uses two disparate distribution schemes: data distribution; and control distribution. In the rst case, cooperation between processes is achieved by sending messages, whereas shared variables are applied in the second case. This latter technique is made possible by implementing a shared virtual memory as part of the operating system. We present brieey an example of such a mechanism, called KOAN, which has been designed speciically for the iPSC/2 hypercube. This shared virtual memory allows us to compare experimentally the performances of the two distribution schemes. Theoretical execution models are also addressed in order to extrapolate expected performance on a large number of processors. R esum e Ce papier pr esente des r esultats d'exp erimentation d'un dispositif de m emoire virtuelle partag ee sur une architecture a m emoire distribu ee (l'hypercube iPSC/2). Deux versions parall eles de l'algorithme de Gram-Schmidt modii e sont pr esent ees. La premi ere est fond ee sur une parall elisation orient ee par les donn ees o u la coop eration entre processus est r ealis ee a l'aide d` echange de messages. La seconde exploite une parall elisation orient ee par le contr^ ole o u les pro-cessus communiquent par variables partag ees. Cette deuxi eme technique est rendue possible gr^ ace a la mise en uvre d'un dispositif de m emoire virtuelle partag ee. Nous pr esentons bri evement un exemple d'un tel dispositif, appel e KOAN, qui a et e r ealis e pour un hypercube iPSC/2. Ce dispositif nous permet ainsi de comparer les deux sch emas de parall elisation. Des mod eles d'ex ecution sont pr esent es aan d'extrapoler les performances de ces algorithmes sur un nombre important de processeurs.