__copy:: H->D: failed: invalid argument

[nikallikou]

Hello,

I installed without problems RapidCFD on Ubuntu 18.04, cuda 9.2.
GPU is Quadro K4200. On a simpleFoam case of a simple cylinder, (inlet with fixed velocity, outlet with fixed pressure) and number of cells 1.7M, rapidcfd works fine. The memory usage at this case is about 50% (2G out of 4). The identical case where i increase to 2M i get the following error:

Create time

Create mesh for time = 0

Reading field p

Reading field U

Reading/calculating face flux field phi

terminate called after throwing an instance of 'thrust::system::system_error'
what(): __copy:: H->D: failed: invalid argument

It seems like a memory issue to me although i hardly believe that the addition of 0.5 M cells make the memory usage from 2G to more than the limit of the GPU.

(I also tried to install another GPU dedicated for the video output and left as the only device for cuda the QK4200 GPU. Rapidcfd works again for the 1.5M but still the same error at 2M)

Do you have any idea?

Thank you very much for your help,
Nikos

[TonkomoLLC]

Hello,

Please also see my reply to issue #57 -

I don't have all the equipment setup right now in a way to confirm that this is a memory issue, as you suspect.

As I wrote about in #57, I have a large damBreak test case that runs fine with 2x K20 GPU's (Ubuntu 16.04, CUDA 8.0), gives a memory error on a single K20 GPU, and a __copy:: H->D error on a machine with Ubuntu 18.04 + CUDA 9.0 + a Titan Black GPU. Both a single GPU and a Titan Black GPU have too little memory to run this case. Therefore, like you, I am suspicious that you are getting a memory error.

This said, I do not know why a modest increase in grid size is doubling your memory usage.

However, since you do have two GPU's, you could try running two GPU's in parallel, as I discussed in issue #57

I have been trying to support CUDA 9 with RapidCFD, since it's so convenient to use CUDA 9 with Ubuntu 18.04. This said, I tend to use Ubuntu 16.04 and CUDA 8.0 with K20 GPU's on larger cases. I know this is an older configuration, but it's been working well for me with RapidCFD.

Good luck with your troubleshooting.

Best,

Eric

[manasi-t24]

Hi Nikos,

I see that you have run the simpleFoam solver for a simple cylinder test case with 1.7 M cells successfully using RapidCFD.
I am trying to compare the performance of RapidCFD with OpenFOAM on CPU's and for that I need different test cases. Would you be kind enough to provide me with your simple cylinder test case files?
Thanks and Regards,
Manasi