PRACTICAL ASPECTS AND EXPERIENCES - SCALABILITY AND LOAD BALANCING INADAPTIVE ALGORITHMS FOR MULTIDIMENSIONAL INTEGRATION

A parallel adaptive algorithm for the approximate computation of a mul ti-dimensional integral over an hyperrectangular region is described. This is a more general version of the well known algorithms presented in Gent (1987) and Lapegna and D'Alessio (1993) and it has been develo ped for an efficient implementation on a MIMD distributed memory multi processor. In order to achieve a good scalability, all global communic ations have been removed from the algorithm. The processor's network h as been configured as a multidimensional periodical mesh, to distribut e information about the integrand behavior fast enough in order to ach ieve a good load balancing. Test results on the Intel Touchstone Delta System are given.