[mvapich-discuss] How to lower MPI standard?

[Åke Sandgren]

On 12/02/2014 11:13 PM, Kin Fai Tse wrote:
> Dear all,
>
> I am running VASP 5.3.5 with MVAPICH2.1a and encounter this error for
> some of the runs:
>
> aborting job:
> Fatal error in PMPI_Allgatherv:
> Invalid buffer pointer, error stack:
> PMPI_Allgatherv(1212): MPI_Allgatherv(sbuf=0x9da0cd0, scount=1944,
> MPI_DOUBLE_COMPLEX, rbuf=0x9d8a050, rcounts=0x97e5710, displs=0x97e57b0,
> MPI_DOUBLE_COMPLEX, comm=0x84000003) failed
> PMPI_Allgatherv(1160): Buffers must not be aliased
>
> Googling suggest that this is because the code is not MPI 2.2 standards
> compliance.
>
> Here is the link to Peter Larsson's blog for the detail:
> https://www.nsc.liu.se/~pla/blog/2014/04/30/vasp535/
>
> Is there any setting similar to I_MPI_COMPATIBILITY=4 in MVAPICH2?

Fix the VASP code instead.
--- a/dfast.F
+++ b/dfast.F
@@ -325,7 +325,7 @@ MODULE dfast
           CALL GGEMM( TRANSA, TRANSB, N1,  PGEMM_HANDLE%NCOL( 
PGEMM_HANDLE%COMM%
                B(1,1+PGEMM_HANDLE%OFFSET( PGEMM_HANDLE%COMM%NODE_ME)), 
LDB, BETA

-         CALL MPI_allgatherv 
(C(1,1+PGEMM_HANDLE%OFFSET(PGEMM_HANDLE%COMM%NODE_
+         CALL MPI_allgatherv (MPI_IN_PLACE, 0, MPI_DATATYPE_NULL, &
                C, PGEMM_HANDLE%NCTOT, PGEMM_HANDLE%OFFDATA, MPIDATA, 
PGEMM_HANDL
        ELSE
           WRITE(*,*) 'internal error in PARALLEL_GGEMM: the second 
matrix needs


-- 
Ake Sandgren, HPC2N, Umea University, S-90187 Umea, Sweden
Internet: ake at hpc2n.umu.se   Phone: +46 90 7866134 Fax: +46 90-580 14
Mobile: +46 70 7716134 WWW: http://www.hpc2n.umu.se