ABINIT Discussion Forums

Dear ABINIT experts,

I am trying linear response calculations on an amorphous material with 108 atoms. The package gave MPI_Allreduce errors. In this forum, geomatteo recommended that I recompile the package without MPI_INPLACE. The cluster administrators kindly reconfigured the package(ABINIT-7.6.2) for me. But now I get the following message. The ABINIT package itself seems to abort the job. Could you help me with the solution? I have tried with 8, 60, 150 and 100 processors with a wall time of 24 hours. The job was best with 60 processors.



PSIlogger: Child with rank 62 exited with status 13.

application called MPI_Abort(MPI_COMM_WORLD, 13) - process 62Fatal error in MPI_
Allreduce: Other MPI error, error stack:
MPI_Allreduce(855).......: MPI_Allreduce(sbuf=0x6a96630, rbuf=0x6882010, count=6
84, MPI_DOUBLE_PRECISION, MPI_SUM, MPI_COMM_WORLD) failed
MPIR_Allreduce_impl(712).:
MPIR_Allreduce_intra(507):
mpid_irecv_done(98)......: read from socket failed - request state:recv(pde)done
MPIR_Allreduce_intra(442):


Sincerely,
Uma

Could you please give your input file ?

Dear Delaveau,

Thank you for offering to help. But now my problem is solved. I was able to run the code with ABINIT7.6.2 (compiled without mpi-in-place), using 75 processors. The linear response calculations went smooth. But due to want of wall time, the job stopped which I continued as another job. However now when I use the mrgddb code, I get the error,

Comparing integers for variable ngfft.
Value from input DDB is 96 and
from transfer DDB is 90.
Action : check your DDBs.

I need to find a way out of this. I wish I don't have to run the response function calculations again..

Sincerely,
Uma

Got the solution from an earlier post. Just manually change the ngfft in the DDB's to match. Then mrgddb runs smooth.
Thanks.
Uma