Computational fluid dynamics simulations sometimes beat experiments

Mannheim, 21-6-97 Numerical simulations of unsteady flows in for instance a tube or around an aeroplane, are problems that require huge amounts of computing time. Egon Krause from RWTH Aachen, explained this at the Mannheim Supercomputer conference. In one simulation, unexpected eddies popped up, that were later confirmed by experiments. Krause uses a large SNI-VPP30 supercomputer for his simulations.

Krause and his group study external and internal flow problems as for instance a steady hypersonic flow around a model for the first stage of a space transportation system. The ELAC 1 is a model for a two-stage-to-orbit space transportation system. Calculating the plume jet required a CPU time of 40 hours and 200 Mbyte of memory on a single processor VPP300.

Turbulence is one of the most difficult features of flow to simulate. In contrast to statistical turbulence models, which are based on time averaging, large-eddy simulations are solutions to the Navier-Stokes equations, a complex set of equations exactly describing all the details of the flow. It took 400 hours of CPU time on a 4 processor VPP300 to solve this problem.

Parallelization of the algorithms involved in this flow simulations can, according to Krause, be done with 70% efficiency. On a computing centre machine parallel production jobs are not always an advantage, since the number of available processors is often small compared to the total number of jobs, so that the overall throughput is reduced through parallelization. On the other hand, parallelization is required if the memory available on one processor is not sufficient for a problem.

Krause expects parallelization to become even more important for his kind of applications in the near future when massively parallel machines with powerful vector processors are installed more often.

Ad Emmen