[CP2K-user] Are there any clear instructions for building cuda-enabled cp2k with MPI that work?

[Travis]

Did you build OpenMPI to be CUDA-aware? I don't know if things changed from 
OpenMPI 2.X to 3.X but it used to be something you specified at the 
configure stage.

https://www.open-mpi.org/faq/?category=buildcuda

Your boost is limited to ([#threads - #gpus] + [#gpus*#speedup]) / 
#threads, so it may be worth asking whether it is worth sinking so much 
time into. Adding a single P100 (~5 TFLOPs in double precision) to a Xeon 
4108 (16 threads, ~0.2 TFLOPS) would give about 2.5x the performance, while 
jumping to a Xeon 8180 (56 threads, ~2 TFLOPS) would give a boost of less 
than 3%.

-T


On Friday, May 17, 2019 at 4:08:51 PM UTC-3, David Gunter wrote:
>
> I have also attempted the same build process using only GCC-7.4.0 with no 
> MKL. I get the same segfault behavior
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20190517/30b246e3/attachment.htm>