Adaptive domain decomposition and parallel cfd

Publisher Summary This chapter discusses a procedure for distributed flow field calculations based on the coarse grain parallelism inherent in domain decomposition methods. The novelty of the method is driven by the combination of three circumstances: the availability of a cluster of high-end workstations with aggregate computing power in excess of mainframe supercomputers; the continued research and interest in fast, highly implicit flow solvers; and the existence of the modular system for computational fluid dynamics (CFD), which is uniquely constructed to facilitate the application of these solvers to complex multizone geometries. In its most popular incarnation, parallelism based on domain decomposition is concerned with mapping subdomains onto processors, balancing loads, minimizing the transfer of boundary data, and synchronizing communications. These concerns arise naturally in an environment where a fixed collection of processors is dedicated to a single calculation and where the minimization of elapsed time is the primary objective. In such an environment, it follows that the most efficient calculation is one in which the available processors are kept constantly busy. The chapter explores the behavior of an essentially chaotic approach in which the sequence of calculations is determined dynamically as the results unfold. In this approach, subdomains are formed adaptively in response to the residual field of the evolving solution. These subdomains are assigned to processors as the need arises. There is no fixed association between processors and subdomains; the subdomains themselves are not permanent entities. There is cost associated with this approach, but, once accepted, it is mitigated by a pronounced increase in flexibility.