parallelization issue, EPW crashes at start

[MaxS]

Hello, the system of interest has a 4x4x4 k-mesh and 27 irreducible k-points. EPW runs fine with

Code: Select all

mpirun -np 16 epw.x -npool 16 < epw.in

and also with 4, 8 and 12 for np/npool. The problem occurs when I specify 27 for np/npool. Then EPW crashes with malloc(): memory corruption right after printing

Code: Select all

     G-vector sticks info
     --------------------
     sticks:   dense  smooth     PW     G-vecs:    dense   smooth      PW
     Sum       13239    5287   1517               919719   232591   35663

It also crashes for 15, 20 and 24. I don't understand this behavior. I tried to run nscf before EPW with 16 and 27 for np but it had no influence on the problem.

However, the calculation epw1.in in QE/test-suite/epw_base runs fine with

Code: Select all

mpirun -np 27 epw.x -npool 27 < epw1.in

(the system there also has 27 irreducible k-points).

What can be the cause for this behavior?

[sponce]

Dear MaxS,

This is a surprising behavior indeed.

Could you try the following:
- use the latest EPW v 5.1 that was release last week
- try all the epw tests from the q-e/test-suite and see if you can find one that reproduce your issue.
- try making your system cheaper for testing (lower the ecut drastically).
- try changing the psps

If that does not work, you can send me your inputs (scf, nscf, ph, epw) and psp and I will take a look.

Best wishes,
Samuel

[MaxS]

Dear Samuel,

I compiled QE with the gnu compiler v8.2 and with openmpi v3.1. The problem has not occurred ever since. It looks like several runtime issues appear when QE is compiled with the intel compiler. I will stick with the gnu compiler form now on.

Thanks for your suggestions and best regards,
Maxim Skripnik