[gmx-developers] Gromacs 2018 detects NVML but doesn't link to it unless set explicitly

[Jernej Zidar]

Hi,

I want to report a bug that I encountered when compiling the final version
of Gromacs 2018.

The cmake command I issue is this:

CMAKE_PREFIX_PATH=/home/zidar/utils/fftw-3.3.6-gcc/ cmake ../gromacs-2018
-DGMX_GPU=on -DGMX_MPI=off -DGMX_FFT_LIBRARY=fftw3
-DCMAKE_INSTALL_PREFIX=/home/zidar/utils/gromacs-2018-gcc
-DGMX_SIMD=AVX2_256

And the NVML library is found:

- Found NVML: /usr/lib/x86_64-linux-gnu/libnvidia-ml.so

When I compile mdrun, the process completes without a problem, yet the
support is not there with mdrun complaining mdrun was compiled without NVML.

After the checking the various cmake-generated files, it appears the
problem is in the CMakeCache.txt file, where the relevant options is set to
OFF (line 698) despite being detected and presumably enabled.

//Use NVML support for better CUDA performance

GMX_USE_NVML:BOOL=OFF

After manually editing the file, the support is there.

I was able to replicate this behaviour on two different workstations with
the following configurations:
- Debian, GCC 6.3, CUDA 9.1
- Ubuntu, GCC 6.4, CUDA 9.0

I've also found that setting the NVML support on explicity using the
switch "-DGMX_USE_NVML=ON" makes everything work as advertised. Perhaps
this behaviour should be mentioned in the documentation.

Cheers,
Jernej
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20180115/e6415fc5/attachment.html>