Run-time partitioning of scientific continuum calculations running on multiprocessors
Scott B. Baden · eScholarship (California Digital Library) · 1987
A wide range of scientific continuum calculations typically concentrate cornputational effort non-uniformly over localized regions of physical space.We present a run-time partitioning strategy, intended for such methods, that distributes work evenly across a team of processors and that can exploit the spatial localization present in the original computation in order to avoid high overhead costs.We tried out our strategy on Anderson's Method of Local Corrections, a type of vortex method for computational fluid dynamics.Because computational effort follows particles that congregate and disperse irregularly about the domain, this problem is hard to partition in a way that distributes the work evenly among the processors.We ran experiments on 32 processors of an Intel Personal Scientific Computer -a message-passing hypercube multiprocessor -and on 4 processors of a Cray X-MP -a ~hared-memory vector architecture -and achieved good parallel speedups of 22 and 3.6, respectively.The partitioner may I feel privileged to have worked with my advisor, W. Kahan.True to his word, Vel will graduate no student before his time; but, the rewards of meeting Vel's challenges will be with me always.A special word of thanks goes to Phil Colella, who was ~y fourth committee member.The distinction as fourth member is exceptional, but so is Phil; he made a special effort to show me the wonders of what had once seemed the dark and mysterious side of computer science -computational mathematical physics -while asking very little in return.I have fond memories of our lively discussions at the Cafe Roma, where our friendship was kindled.And yes, many thanks go to Phil's wife,