VASP support site: Forums / Installation problems / Problems with MPI vasp at runtime

VASP support site

Main Menu

Home
Links
Forum
Members
Stats
Upload
How to report bugs

Welcome

Online

Guests: 8
Members: 0
On this page: 1
Members: 6707, newest: jer7oll

Forums

VASP support site -> Forums -> Installation problems

<< Previous thread | Next thread >>

Problems with MPI vasp at runtime

Moderators: admin.

Author

Post

brockp

Thu Jun 08 2006, 07:45PM

Registered Member #712
Joined Thu Jun 08 2006, 07:33PM
posts 4

I am a sysadmin helping a user install vasp on our linux (RHEL 4.0) opteron cluster. The compiler is PGIF90 6.1 and the MPI lib is OpenMPI 1.0.2 the serial version builds and runs just fine but the paralell version gives the following error,

running on 2 nodes
[nyx.engin.umich.edu:28430] *** An error occurred in MPI_Cart_create
[nyx.engin.umich.edu:28430] *** on communicator MPI_COMM_WORLD
[nyx.engin.umich.edu:28430] *** MPI_ERR_OTHER: known error not in list
[nyx.engin.umich.edu:28430] *** MPI_ERRORS_ARE_FATAL (goodbye)
[nyx.engin.umich.edu:28431] *** An error occurred in MPI_Cart_create
[nyx.engin.umich.edu:28431] *** on communicator MPI_COMM_WORLD
[nyx.engin.umich.edu:28431] *** MPI_ERR_OTHER: known error not in list
[nyx.engin.umich.edu:28431] *** MPI_ERRORS_ARE_FATAL (goodbye)
1 additional process aborted (not shown)

This is a regular OMPI error, and i have contacted the devs of openmpi, i am posting here to see if this is a problem anyone else has seen and if so how/if they were able to fix this problem.

Brock

brockp

Fri Jun 09 2006, 02:03PM

Registered Member #712
Joined: Thu Jun 08 2006, 07:33PM
posts 4

Looks like the problem isnt with openMPI here is the result rebuilding everythign with lam-7.1.2

bash-3.00$ mpirun -np 2 ./vasp
running on 2 nodes
MPI_Cart_create: invalid dimension argument: Invalid argument (rank 0, MPI_COMM_WORLD)
Rank (0, MPI_COMM_WORLD): Call stack within LAM:
Rank (0, MPI_COMM_WORLD): - MPI_Cart_create()
Rank (0, MPI_COMM_WORLD): - main()
MPI_Cart_create: invalid dimension argument: Invalid argument (rank 1, MPI_COMM_WORLD)
Rank (1, MPI_COMM_WORLD): Call stack within LAM:
Rank (1, MPI_COMM_WORLD): - MPI_Cart_create()
Rank (1, MPI_COMM_WORLD): - main()

Could it be a problem with the users input that the problem cant be broken down correctly ? This input works fine on the serial version of vasp

job

Wed Jun 14 2006, 11:03AM

Registered Member #371
Joined: Tue Aug 16 2005, 07:44AM
posts 55

Have you compiled vasp with -i8 and the mpi library with default settings? That won't work.

c00jsh00

Thu Jun 29 2006, 05:22AM

Registered Member #477
Joined: Tue Nov 15 2005, 09:01AM
posts 7

Hi,

We have the similar problem, have you solved the problem yet?

admin

Mon Jul 24 2006, 11:20AM

posts 2344

please check if the LAM was compiled in the same bit-mode as you used for the compilation of vasp

brockp

Tue Sep 26 2006, 07:02PM

Registered Member #712
Joined: Thu Jun 08 2006, 07:33PM
posts 4

job wrote: ...
Have you compiled vasp with -i8 and the mpi library with default settings? That won't work.

I tried rebulding the OpenMPI with -i8 but chokes on configure ( i have contacted the devs)

I also tried forcing 4 byte integers -i4 and rebuilt both vasp.4.lib and vasp.4.6 and i get a new error:

[brockp@nyx VASP.run]$ mpirun -np 2 -v ./vasp
running on 2 nodes
[nyx.engin.umich.edu:31483] *** An error occurred in MPI_Cartdim_get
[nyx.engin.umich.edu:31483] *** on communicator MPI_COMM_WORLD
[nyx.engin.umich.edu:31483] *** MPI_ERR_COMM: invalid communicator
[nyx.engin.umich.edu:31483] *** MPI_ERRORS_ARE_FATAL (goodbye)
distr: one band on 1 nodes, 1 groups
1 process killed (possibly by Open MPI)

So i still have not made any progress. I also added the -Ddebug to the flags, but vasp did not display anything.

Also what does -Dkind8 mean?

brockp

Tue Oct 03 2006, 01:07PM

Registered Member #712
Joined: Thu Jun 08 2006, 07:33PM
posts 4

The problem was solved using the following:

lam-7.1.2
Open MPI would not work with vasp this is unfortonate, both mpich and lam are nolonger dev. Moving to more uptodate MPI libs like OpenMPI would be a plus in the future. Im not sure if its OpenMPI or VASP causing the problem so i will pass it on to the OMPI devs see if we can fix it.

PGI 6.1 -i4
Matchin the size of LOGICALS and such was a real pain, Its not documented anyware but the default PGI make file for linux has -i8 in the Makefiles. This caused quite a headache. This MUST match what your MPI lib was built with.

VASP is running now int MPI GoTO was very slow on the example case i had (dont know why) ACML3.5 was slightly faster than ATLAS.

This was on OPT 244 with GIG-E non blocking + Jumbo frames networking. Hope this helps anyone else. If you want I can provide Makefiles for anyone having trouble.

Brock
1(734)936-1985
Center for Advanced Computing
University of Michigan (Ann Arbor)

atogo

Tue Oct 17 2006, 06:06AM

Registered Member #111
Joined: Tue Mar 22 2005, 07:45AM
posts 10

I met similar problem. In my case, I could solve it by following:

Set the compiler path directory like as,
FC=/usr/foo/bar/bin/mpif90
'FC=mpif90' with $PATH doesn't work. A shared library is missing. I don't know why.

Another choice is to link staticaly ( libmpi.a, liborte.a and libopal.a in openmpi case). Remenber to copy header files from 'include' directory in openmpi.

Date / Time

Counter

Today: 393 (unique:122)
Ever: 1095185192 (unique:1093721993)
Record: 485092 (unique:485090)