Parallel CFD Computation on Unstructured Grids

Abstract

This chapter discusses partitioning issues of unstructured grids and describes the message communication together with the data structure. The issues involved in parallelizing cell-centered finite volume schemes for solving the Reynolds-averaged Navier–Stokes (RANS) equations on triangular unstructured grids on a workstation cluster—a parallel computer system based on the workstation networks under the parallel environment supported by parallel virtual machine (PVM) software—are explained. The recursive spectral bisection (RSB) and multilevel graph partitioning (MGP) methods are applied and a method of construction the message passing relations and data structures is proposed. The definition of three types of elements is given and illustrated with example. Performance of parallel computing on airfoil problems has been demonstrated. Message passing, data structure (sending and receiving) and communication graphs have been illustrated and efficient computation has been achieved. Some issues connected with parallelizing a flow solver are domain decomposition, message communication, data structure, and parallel algorithms. The partitioning method of unstructured grids for parallel computing has been investigated by many researchers. The methods can be broadly classified into geometry-based and graph-based algorithms. It appears that the graph-based algorithms, in particular, the recursive spectral bisection (RSB) technique, and the multilevel graph partitioning (MGP) produce much better results. Messages exchanged between neighboring sub-domains should be done efficiently during the iteration at the end of which the code should be parallelized.