Multiprocessing a sparse matrix code on the Alliant FX/8

摘要

We have recently multiprocessed a code for the direct solution of sparse linear equations on the Alliant FX/8. We discuss several issues which are involved, all of which are of relevance to any shared memory multiprocessor. Among these issues are the dynamic allocation of data, the management of task queues, task spawning, and the effect of controlling the granularity. We also show runs of our code under the SCHEDULE package from Argonne which presents a portable interface to users of parallel machines, allows the user to define the computational graph, and has very useful graphic output to a SUN workstation. Our tailored code attains a speedup by a factor of about six on the eight processors of the Alliant. We suggest ways of improving it further.