Compiling CP2K 6.1 with OpenMPI 3.1.2

Matthias Krack matthia... at psi.ch
Fri Sep 21 14:57:02 CEST 2018


Dear Margaret

I have run the Linux-x86-64-gfortran.popt regression test using gcc-7.3.0 
and openmpi-3.1.2 on a local cluster node. Only 8 out of 3038 regression 
tests failed and the timings are reasonable. Note, that the regression test 
within a node may run very slow with OpenMPI when all the parallel 
regression tests are launch with the default mpiexec/mpirun setting 
"--bind-to slot". That default may cause that all jobs will use the same 
pair of CPU cores which will result in an extreme overloading of these 
cores. You can overwrite that default with the flag "--bind-to none". On 
the other hand, the default setting is better for production runs when one 
CP2K job populates all cores of several compute nodes, since slot/core 
binding is beneficial for such cases.
The OpenMPI 3.1.2 release is quite recent (Aug 22). If possible, I 
recommend to use an alternative MPI implementation like MPICH for CP2K 
which is potentially less troublesome for the time being.

HTH

Matthias

On Wednesday, 19 September 2018 01:46:14 UTC+2, Margaret Davis wrote:
>
> Hello,
> Has anyone tried to compile and run regression tests for CP2K 6.1 with 
> OpenMPI 3.1.2? I've compiled CP2K 6.1 a number of times successfully with 
> other versions of OpenMPI. I'm trying to compile it now and I think the 
> OpenMPI version is the only thing that has changed. My regression tests are 
> running much slower than they have in the past and certains jobs within the 
> regressions tests are being killed. I attached my error summary. 
> If anyone has experienced this problem or knows how to fix it that would 
> be awesome.
> Thanks,
> Margaret Davis, Syracuse University 
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20180921/2a507840/attachment.html>


More information about the CP2K-user mailing list