A Data Structure for Parallel L/U Decomposition

Some new results are presented concerning the pivoting of large systems of linear equations with respect to parallel processing techniques. It will be assumed that the processing of a pivot takes one time slot. The pivoting problem is studied by means of an associated graph model. Given a triangulated graph a set of label classes is established. Class k contains all pivots which may be processed in parallel during the kth time slot. The label classes are used to establish the elimination-tree (e-tree). The e-tree is a spanning tree for the given graph. The critical path in the e-tree indicates the minimum number of time slots necessary to complete the L/U-decomposition. Furthermore, the earliest and latest admissible time slot for the processing of every pivot may be derived, such that the critical path is not affected. The e-tree can be seen as a data structure to guide parallel processing based on sparsity.