Microtasking versus Message Passing Parallelisation of the 3D-Combustion Code AIOLOS on the NEC SX-4

Abstract

The chapter is divided into four main sections that include introduction of the mathematical modeling of industrial furnaces using the 3D-combustion code AIOLOS, review of characteristics of different parallel programming models, and presentation and discussion of results of the computational performance of these models. The 3D-combustion code AIOLOS is a tool for the mathematical description of turbulent combustion and pollutant formation in industrial furnaces. The chapter shows that the high computational resources required for accurate predictions can be met by the utilization of parallel vector computers like the NEC SX-4. The modeling of turbulent combustion and pollutant formation processes in industrial scale furnaces is based on the numerical solution of the balances of mass, energy, species concentration and momentum. This leads to approximately 50 transport equations which have to be solved. Parallel programming models can be classified according to the way in which data management and inter-processor communication are realized. Three models have received intensive attention in the past years: a coarse grain, distributed memory model; a fine grain, distributed memory model; and a fine grain, shared memory model. To exploit the potential of parallel execution AIOLOS has been parallelized in the past with these three programming models. A detailed description of the implementation as well as a discussion of performance results on different parallel platforms is provided. The comparison of a shared memory programming model (Microtasking) and a distributed memory programming model (MPI) on a shared memory system has been presented. The hardware used for this comparison is a parallel vector computer (NEC SX-4/32). Both programming models showed an acceptable parallel performance reducing the total computing time required for full scale furnace simulations. The basic idea of this model is to assign a number of domains to different shared memory nodes connected in a distributed memory manner.