Flotilla Demo 2.07

[Leo Lazauskas]

Originally Posted by jtremolo

Currently we are using it for rotorcraft wake simulations. We have run hundreds of test cases and all matched up with experimental data and current commercially available CFD solutions. As far as fast goes, when we ran a turbulent rotor simulation on the CPU it took approximately 8 hours on a 128x128x128 grid. When we ran the same problem using CUDA it took 11 minutes.

That's about 50 times faster which is very impressive.

A "typical" run for a RANS simulation for one speed of an Olympic rowing shell (according to Mola's recent PhD thesis) takes 24 hours on a cluster of 8 Linux workstations. Using CUDA might reduce that to about 30 minutes.

Originally Posted by jtremolo

As far as running your examples, I can definitely get the skin friction drag on a flat plate. The wave forms will take a little bit of playing, I'm used to aircraft not boats. However, I am definitely up for trying it and seeing what happens.

I'd love to see your estimates for the skin-friction at Rn of 10^6, 10^7, 10^8, 10^9, and 10^10.

Forget about the complete wave patterns. But the drag of the NPL hull series in the pdf documents I attached earlier would give your code a real challenge.
Incidentally, a 50 times speed-up using GPUs would mean Flotilla could complete the entire set in a couple of minutes, which is definitely a worthwhile improvement!

Another good test of your code would be to estimate the total resistance and squat of the current "international benchmark" hull, the DTMB 5415 destroyer hull. Experiments are being conducted on this hull at about 40 towing tanks around the world so there will be plenty of experimental data to use for comparisons.

Originally Posted by jtremolo

If you are interested, I can send you the thesis my research partner wrote when he graduated. It talks about the whole process we did and has some pretty pictures

Sure, send it to lazauskas@gmail.com

I have no doubts that there are pretty pictures. They don't say CFD stands for "Colour For Dollars" for nothing

Have fun!
Leo.

[DHMzip]

I am a yacht design student, and last year we a tank test on an NPL series hull. i would be very interested to see the results of this CFD done with CUDA. I have been thinking about getting an Nvidia TESLA card for some time now (not only to do CFD, but also for Rendering).

I would also be interested to see that paper. my email is: davidhmassey@gmail.com

jtremolo- how do you find CUDA? is the card worth it compared to building a small computer cluster?

thanks,
David

[DHMzip]

PS.

thanks Leo for your reports. this is very interesting stuff.

~David

[Leo Lazauskas]

Originally Posted by DHMzip

PS.

thanks Leo for your reports. this is very interesting stuff.

~David

My pleasure.

Your real thanks should go to Patrick Couser and his colleagues at Uni Southampton for the splendid set of experimental data.

Leo.

[jtremolo]

David,
I would say for calculations CUDA is the way to go. I think the results speak for itself, We dropped our computation time from 8 hrs to 11 min with a full 3D Navier-Stokes solver. The card we are using is the GTX 470 which is approx $400. I don't know of any computer cluster that can get that kind of speed up for that money. Also, due to the architecture of the CUDA system it is almost infinitely scalable dependent on the number and size of cards.

If you want more information on CUDA go to http://nvidia.com/cuda. If you buy a card make sure it is CUDA capable, some of the older ones aren't. There is a list of CUDA capable cards on NVIDIA's site.

[Leo Lazauskas]

Originally Posted by jtremolo

David,
I would say for calculations CUDA is the way to go. I think the results speak for itself, We dropped our computation time from 8 hrs to 11 min with a full 3D Navier-Stokes solver. The card we are using is the GTX 470 which is approx $400. I don't know of any computer cluster that can get that kind of speed up for that money. Also, due to the architecture of the CUDA system it is almost infinitely scalable dependent on the number and size of cards.

If you want more information on CUDA go to http://nvidia.com/cuda. If you buy a card make sure it is CUDA capable, some of the older ones aren't. There is a list of CUDA capable cards on NVIDIA's site.

Thanks, you're a wealth of information!

I'll be interested to see how you set up the flow-field in the region of a transom stern using your "full 3D Navier-Stokes solver". It might be a bit more complicated than you imagine if you have not used it for free-surface flows before.

Cheers,
Leo.

[jtremolo]

Leo,
Calculating the hull is definitely going to be a challenge, like I said earlier i'm used to things moving through air. So, I don't really have much experience with free surface wave boundary conditions etc. However, hull design is definitely an interest of mine, so this would be a neat tool to have. I am working on getting a first test run ready but, it may take me a little bit to dial in my program for use on hulls.

[Leo Lazauskas]

Originally Posted by jtremolo

Leo,
Calculating the hull is definitely going to be a challenge, like I said earlier i'm used to things moving through air. So, I don't really have much experience with free surface wave boundary conditions etc. However, hull design is definitely an interest of mine, so this would be a neat tool to have. I am working on getting a first test run ready but, it may take me a little bit to dial in my program for use on hulls.

If you are genuinely interested in the role of CFD in ship hydrodynamics, keep an eye on the participants' submissions to the Gothenburg 2010 Conference that will be held in Dec. 2010.

http://www.gothenburg2010.org/

Leo.