Page 1 of 1

VASP crashes with intel mpi for more than 1 node

Posted: Thu Jun 06, 2013 5:00 pm
by Wasabi
Dear VASP Forum,


I'm trying to use vasp.5.3.3.

I compiled using intel MPI (IMPI) and MKL without any problems. Here are some important lines from my Makefile:

Code: Select all

??
FC=mpiifort??
FFLAGS?=??-FR?-names?lowercase?-assume?byterecl

OFLAG=-O2?-ip?-xHost

BLAS=?-Wl,--start-group?$(MKL_PATH)/libmkl_intel_lp64.a?$(MKL_PATH)
/libmkl_sequential.a?$(MKL_PATH)/libmkl_core.a??-Wl,--end-group?-liomp5?-lpthread?-lm

LAPACK=?$(MKL_PATH)/libmkl_intel_lp64.a


Our machine is a Linux cluster with 16 cores per node. Running up to 16 cores (1 node) we do not experience any problems.

Running on more than 16 cores with a simple (60 metal atoms) test systems seems to be perfectly smooth.

However, when I go to a larger system (~200 atoms) I get this error when I use more than one node:

Code: Select all

Fatal?error?in?PMPI_Alltoall:?Pending?request?(no?error),?error?stack:
PMPI_Alltoall(990)......:?MPI_Alltoall(sbuf=0xadc2b20,?scount=226,?MPI_DOUBLE_PRECISION,?rbuf=0xa805b00,?rcount=226,?MPI_DOUBLE_PRECISION,?comm=0x84000005)?failed
MPIR_Alltoall_impl(800).:?
MPIR_Alltoall(761)......:?
MPIR_Alltoall_intra(567):?



Do you recognize this error, what could be the reason for it?
If it helps, I can add the "traceback" option for the compilation ...


I have "ulimit -s unlimited" in my submission script ...

Also worth mentioning: Using more than 1 node I always get this error:

Code: Select all

[
forrtl: File exists
forrtl: severe (10): cannot overwrite existing file, unit 12, file /home/.../WAVECAR
/code] 

Of course I can circumvent this issue with "touch WAVECAR" and "touch CHGCAR" in my submission script. But maybe it is related to my problems?

ANY help is very much appreciated.

w.

VASP crashes with intel mpi for more than 1 node

Posted: Wed Jun 12, 2013 4:52 pm
by admin
This error seems to be an bug in the MPI installation, please (MPIR_Alltoall), please ask your sys admin to check it