A hierarchical parallelization concept for a high-performance Navier-Stokes solver

Thomas Hauser, G. Xiong, and George Huang
Department of Mechanical Engineering
University of Kentucky
Lexington, KY 40506

Abstract

The future of high-performance computer hardware will continue to evolve in the direction of clusters of SMP computers. In this model, SMP computing nodes are interconnected by fast, high-speed data links. While OpenMP provides a convenient way for shared memory parallel programming, MPI is the natural approach for distributed computing. In this paper we therefore present a hybrid MPI/OpenMP approach in an attempt to combine the best features of the two approaches. The hybrid approach provides the flexibility to choose between shared and distributed memory computing, or a combination of the two.

The CFD code LESTool is used for the simulation of complex geometries by applying the multi-block concept. This strategy is designed to take advantage of the hybrid MPI/OpenMP practice. From the overall pool of processors, different groups of processors are clustered together. These clusters of processors will contain one or more grid blocks and communicate using MPI. Within each cluster the implicit ADI-type solver is parallelized using OpenMP. A complex turbine cascade is used to test the scalability and performance behavior of this approach on a SGI Origin 2000.



Return to High Performance Scientific Computations with Applications main page.