I am researching very large scale CFD applications that can be used in product design.

I am interested in gathering data on the largest CFD simulations that have been run on commercial parallel servers, using commercial CFD codes.

30 million cells, 50 million cells, 100 million cells ?

What is the largest case ?

Are there any highly parallel commercial codes that can handle large cases above 50 million cells, do grid generation for complex geometries automatically and converged simulation - in a few days ?

There was a demo mesh generation code free of charge for a full aircraft configuration posted here sometimes ago, is an aircraft big enough?

Indeed we have clients using Exa's PowerFLOW (highly scalable code using Lattice Boltzmann) who have run cases with over 50 million computational elements. Scalability has been proven on a 128 processor SGI system: super-linear to 32P, linear to 64P, and at 128 performance was about 92%. PowerFLOW's mesh generation is fully automatic. Turnaround times for full size external flow simulations (ground vehicles, complex geometry) can be performed in under 8 hours on highly parallel machines.

Please see www.exa.com for more information. The direct address for the SGI scalability study is:

http://www.exa.com/powerflow/pfvald-scalability.html

Is 50M elements the commercial record as far as anyone knows?

Dave Higbie

At NASA Ames,

They have done 40 million on their special 256 processor Origin 2000, and planned a 60 million case on it. This was using the thin layer Navier Stokes code called Overflow.

Tony

Hello,

The largest I know of is 20 million tet cells of an external flow of a car including wheel arches and braking systems. The model was created in Icem TETRA but I don`t know what it was solved on.

regards

David

http://www.topologies.co.uk

The record as far as I know is 120 million control volumes for a simulation of for a turbulent flow + heat transfer in a nuclear reactor (real geometry, with 7000 rods and rod spacers). Taking 7 variables per cell (u, v, w, p, k, e, T), that't s 870 million degrees of freedom in the FEM jargon.

The mesh generation was not a real problem (repeating pattern), but the simulation could not have been simplified because the boundary conditions were not trivial.

The calculation was done by adapco (Melville NY, www.adapco.com) on a parallel IBM RS 6000 with 122 CPU at the IBM Benchmark and Enablement Centre in Poughkeepsie. The code is the parallel version of STAR-CD and the parallel speedup was impressive.

(NASA, eat your heart out!)

Hrvoje Jasak

in all fairness dave i think you're comparing oranges and tangerines when you compare 50 million lattice boltmann elements to (say) 50 million cells/grid point in standard cfd techniques. by the way is exa good for transonic flows?

Making a direct comparison between lattice-Boltzmann and RANS computational elements is problematic. But 50M is still a lot!

Recently Dimitri in ICASE computed a large problem around an aircraft. You can find that in the ICASE report. He used about 23 million nodes in that computation. Since that is an unsturctured mesh, the Tetrahedra number should be around 140 million. They tried in both SGI and Cray machine.