No activity on GPU when compiled with.

[Former user]

Hello,

I'm under Debian Stretch. I've installed all compilers of 4.9.4 versions in order to be well compiled under cuda.

Here are my commands to compile: sed -i "s/compute_35/compute_35\ -Xcompiler\ -fPIC/g" configure.ac ./autogen.sh make distclean-recursive ./configure --with-cuda --with-fftw --enable-integer-positions

I add -Xcompiler -fPIC in NCCFLAGS in order to compile correctly.

When I launch on default example ( cd examples ; mpiexec ../pkdgrav3_mpi cosmoloy.par ), the program seems to access to GPU (lsof | grep pkdgrav3 | grep nvidia0 returns lots of outputs) but the activity command provided by Nvidia provides nothing.

nvidia-smi --query-compute-apps=pid,process_name,used_memory --format=csv

I try to compile and execute without CUDA support and I got exactly the same execution time... I try to get any GPU activity over 10 different GPUs, I got nothing (p100, k80, k40, gtx1080ti, m2200, etc).

Where is my failure ?

Best regards

[Douglas Potter]

To enable GPU use, you must set the "CUDA Queue Size" on the command line. A suitable value is 8, for example:

mpirun pkdgrav3_mpi -cqs 8 input.par

Yes, I know this is not very intuitive, and a future version will "automatically" enable the GPU by default if it is detected.

[Douglas Potter]

The new version, when compiled with GPU support, will automatically enable the GPU if it is available.