[DFTB-Plus-User] Possible bug in DFTB+ with MPI

[Ben Hourahine]

Hello David,

thoughts so far:

1) Which blas are you using? Sometimes openBlas steals extra threads, is
your top output shown with thread reporting enabled (H)?

2) Are you definitely getting the OMP_NUM_THREADS shell variable to each
of the instances?

3) If this is related to the i-PI mailing list question, is there any
reason not to run 8 serial instances with 1 thread each? The MPI
parallelism overhead on just one process is probably not worth it.

Regards

Ben

On 22/06/2019 14:41, David Furman wrote:
> Hi all,
> I have a question concerning the running of DFTB+ with MPI. There
> seems to be a problem when running more than one DFTB+ instance (with
> MPI parallelization) on the same system.
>
> When I run 4 instances (separate folders) of dftb+ with 2 cores each
> (i.e. 8 cores in total):
> mpirun.openmpi -np 2 dftb+, each %CPU usage drops to 50%.
>
> This is the 'top' output:
> ========================================================
>  35426 user    20   0  423384  40404  18248 R  50.5  0.0   0:37.69 dftb+
>  35434 user    20   0  423080  39720  18248 R  50.2  0.0   0:05.70 dftb+
>  35396 user    20   0  422044  38660  17644 R  49.8  0.0   1:21.35 dftb+
>  35402 user    20   0  423384  40460  18304 R  49.8  0.0   1:09.54 dftb+
>  35403 user    20   0  422044  38872  17848 R  49.8  0.0   1:16.15 dftb+
>  35427 user    20   0  422044  39048  18032 R  49.8  0.0   0:32.93 dftb+
>  35435 user    20   0  421732  38164  17716 R  49.8  0.0   0:09.07 dftb+
>  35395 user    20   0  423384  40360  18204 R  49.5  0.0   1:13.18 dftb+
> =========================================================
>
> Whereas, when I run one instance with 8 cores, the efficiency is 100%
> as expected:
>
> =========================================================
>  35837 user    20   0  441868  35196  19112 R 100.3  0.0   0:04.90 dftb+
>  35828 user    20   0  442580  35676  19120 R 100.0  0.0   0:04.83 dftb+
>  35829 user    20   0  442544  36024  19432 R 100.0  0.0   0:04.91 dftb+
>  35830 user    20   0  441912  35356  19360 R 100.0  0.0   0:04.91 dftb+
>  35831 user    20   0  441688  35048  19304 R 100.0  0.0   0:04.90 dftb+
>  35833 user    20   0  441904  35656  19500 R 100.0  0.0   0:04.91 dftb+
>  35841 user    20   0  441516  34824  19076 R 100.0  0.0   0:04.89 dftb+
>  35843 user    20   0  441224  34456  18868 R  99.7  0.0   0:04.89 dftb+
> =========================================================
>
> I run both cases with OMP_NUM_THREADS=1.
> with open-mpi 1.10.2 and gcc 5.4.0.
>
> Could anyone give a hint about what is wrong?
>
>
> _______________________________________________
> DFTB-Plus-User mailing list
> DFTB-Plus-User at mailman.zfn.uni-bremen.de
> https://mailman.zfn.uni-bremen.de/cgi-bin/mailman/listinfo/dftb-plus-user

-- 
        Dr.  B.  Hourahine,  Senior  lecturer
            SUPA, Department  of  Physics,
             University  of  Strathclyde,
              John  Anderson  Building,
          107 Rottenrow, Glasgow G4 0NG, UK.                    
  +44 141 548 2325, benjamin.hourahine at strath.ac.uk

The  Department is  a partner  in SUPA,  the Scottish
            Universities Physics Alliance

The University  of Strathclyde  is a  charitable body,
       registered in Scotland, number SC015263

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.zfn.uni-bremen.de/pipermail/dftb-plus-user/attachments/20190622/bf90b2fb/attachment-0001.html>