[Beowulf] gpu benchmark

[Michael Di Domenico]

On Wed, Aug 31, 2016 at 10:08 AM, Massimiliano Fatica <mfatica at gmail.com> wrote:
> Point 2 and 3 are easy to fix ( the 1.6.5 binary will run on 1.10.2)
> and you only need the lubcudart.6.5.so ( or make symbolic links from
> the 7.0 to 6.5) in your LD_LIBRARY_PATH.

i'm not sure i agree with that or at least my test didn't

on a test machine i can link the CUDA6 lib, but hpcg looks for
libmpi.so.1 and openmpi 1.10.2 provides libmpi.so.12

thinking i might be able to fake it with a symlink, i can satisfy the
lib dependancy, but then hpcg spits out

Cuda error: generate_problem_cuda in file '../src/CudaKernels.cu' in
line 176: invalid device function .