Synchronization-Free Parallel Implementation of Factored-Implicit and Explicit Techniques

Abstract

This work is motivated by the desire to assess the adaptability of alternating direction implicit and explicit finite difference techniques, used in computationl fluid dynamics, to parallelization on tightly-coupled, shared-memory parallel computers. The aim of the work is to achieve synchronization-free parallel implementations by the introduction and use of load balancing and processor routing schemes. Two representative and popular finite difference techniques, the implicit Beam and Warming and the explicit MacCormack techniques, are parallelized and tested on a 20-processor computer. Several parallel implementations of increasing complexity are investigated in the context of the simple heat and the more computationally intensive curvilinear grid mapping equations. In attaining near-theoretical speedups, the need for processor synchronization has to be eliminated completely. To achieve a synchronization-free parallel implementation, three stratagems are introduced and evaluated. First, each finite diff...