[QE-users] Questions regarding parallelism and memory usage of ph.x calculations
Laurens Siemons
laurenssiemons at hotmail.be
Tue Jun 11 11:53:35 CEST 2019
Dear QE-users,
I tried to run a ph.x calculation on 100 processors of a system with 5 unique k-points so I used the -nk 5 command to divide the calculation into 5 pools of 20 processors. After printing which representations will be calculated in this run, the output returns the following error message:
...
Representation 469 1 modes -A To be done
Representation 470 1 modes -A To be done
Representation 471 1 modes -A To be done
Compute atoms: 149, 150, 151, 152, 153, 154, 155, 156,
157,
===================================================================================
= BAD TERMINATION OF ONE OF YOUR APPLICATION PROCESSES
= PID 16398 RUNNING AT r3c4cn08.hopper
= EXIT CODE: 9
= CLEANING UP REMAINING PROCESSES
= YOU CAN IGNORE THE BELOW CLEANUP MESSAGES
===================================================================================
Intel(R) MPI Library troubleshooting guide:
https://software.intel.com/node/561764
===================================================================================
However, if I perform the same calculation without parallellization over the unique k-points, the calculation achieves convergence without a problem.
Parallellization over the 5 unique k-points works without a problem in the corresponding pw.x calculation. Any idea what can cause this issue?
Except for this, I had an additional question regarding the memory usage of a ph.x calculation. While running such calculations, the memory usage goes up until approximately 200G. I'm performing calculations on a cluster and the amount of memory I can use is limited. Are there general tricks/methods to reduce the memory usage of ph.x calculations?
Thanks in advance,
Laurens Siemons
PhD Student, UAntwerp
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20190611/c7a407bb/attachment.html>
More information about the users
mailing list