Parallelization of Divide-and-Conquer Eigenvector Accumulation

Wilfried N. Gansterer, Joachim Zottl

Abstract: This paper discusses and compares several parallelization strategies for tree-structured computations. In particular, we focus on the parallelization of the eigenvector accumulation process in divide-and-conquer eigensolvers, such as the recently developed block divide-and-conquer (BD&C) eigensolver. We describe a model algorithm for evaluating the performance of several parallel variants of this accumulation process, and we develop a block parallel approach which is shown to achieve good speedup in experiments on PC clusters.