After a mesh is partitioned into subdomains, there remain several operations that have to be performed on the partitions to prepare the input for the parallel program. We refer to these steps collectively as parceling. The steps include generating (i) the communication schedule for each processor, (ii) the global-to-local mapping information, which allows identification of a node or element number on a processor by its global number, and (iii) the nonzero structure of the stiffness matrix on each processor. The last item could be performed in parallel, but it takes little time and provides us with useful statistics on the mesh, so we perform it sequentially. Parceling requires about 2.3 hours and 7.7 Gb memory on the DEC 8400 for the 77 million element San Fernando Basin mesh. The communication graph generated by the parceler is shown in Figure 5.
Figure 5:
Communication graph for the partitioned element mesh
depicted in Figure 4.
Each vertex represents a subdomain and corresponding processor; each edge represents communication between two processors.