DualSPHysics with Quadro 2000 and a Tesla C2075 cards
I just got a new machine with a Quadro 2000 card and a Tesla C2075 card. The system is configured using NVIDIA's maximus utility with the Tesla card being dedicated to computations and the Quadro card dedicated to graphics. However, whenever I run a GPU simulation, DualSPHysics uses the Quadro card not the Tesla card. Does anyone have a suggestion to get DualSPHysics to use the Tesla card?
Thanks.
Daley
Thanks.
Daley
Comments
DualSPHysics has some GPU detection code which parses all available GPUs in the system and then it picks the one it thinks is best, however the decision making process is fairly rudimentary.
You can override the card that it chooses using a simple flag on the command line when you run DualSPHysics. This is described in the manual, however all you need to do is modify the "GPU" flag to include the index of the GPU you wish to use, i.e. if I wanted to use the GPU my system saw as GPU 1 then I would put '-gpu:1'.
That being said, I have also experienced some problems with the Maximum utility, really it is designed to switch for games, which tend to have far more comprehensive code to cover this sort of thing. You may wish to disable that also.
Thanks for your advice.
I have tried altering the code Run_Directory/CASEDAMBREAK/FastCaseDambreak_win64_GPU.bat in the 64-bit Windows OS DualSPHysics from
%dualsphysics% %dirout%/%name% %dirout% -svres -gpu
to
%dualsphysics% %dirout%/%name% %dirout% -svres -gpu:1 (I also tried -gpu:0)
However this did not work.
After some more digging around, I noticed that my system only preforms a GPU simulation when it is configured using NVIDIA's maximus utility with the Tesla card being dedicated to computations and the Quadro card dedicated to graphics and computations.
When I switch it to 'Tesla card dedicated to computations' and the 'Quadro card dedicated to graphics',I get a message saying DualSPHysics is initialising CUDA device and then a window pops up saying 'DualSPHysics_win64.exe has stopped working'. I presume this is because DualSPHysics is trying to use the Quadro 2000 card, but the card is saying it will only do graphics……
After this, removed the Quadro card and ran a GPU simulation with only the Tesla card. I was given the same message about initialising CUDA device.....
At this point I am thinking there is an issue with the Tesla card (either physically or with the driver)? What do you think?
Faulty GPUs can be a nightmare to detect, I recently spent many months trying to find a bug in the code only for it to turn out to be an over-heating GPU!
I would suggest trying to perform some sort of diagnostic on the Tesla.
Tools such as CUDA GPU memtest are a good starting point (http://sourceforge.net/projects/cudagpumemtest/).
I do have to say though I would recommend removing the Maximum utility from your system if you can. It's a nice idea and makes sense on a CAD machine or gaming box, but more manual control is desirable for the numerical analyst.
Personally I operate under Linux almost solely as I find it the faster environment for CUDA crunching and the more productive for software development, so things like Maximus aren't even a consideration.
Let us know how you get on.