Re: [OMPI users] memory leak in alltoallw

[George Bosilca]

Dave,






  Thanks again for your report.
    george.

On Aug 7, 2008, at 1:21 AM, Dave Grote wrote:

Thanks for your report. As you discovered we had a memory leak in the MPI_Alltoallw. A very small one, but it was there. Basically, we didn't release two internal arrays of data-types, used to convert from the Fortran data-types (as supplied by the user) to their C version (as required by the implementation of the alltoallw function). The good news is that this should not a problem anymore. Commit 19314 fix this for the trunk, while commit 19315 fix it for the upcoming 1.3.

Hi,

 Thanks!
    Dave

    program testalltoallw
    real(kind=8):: phi(-3:3200+3)
    real(kind=8):: phi2(-3:3200+3)
    integer(4):: izproc,ii
    integer(4):: nzprocs
    integer(4):: zrecvtypes(0:3),zsendtypes(0:3)
    integer(4):: zsendcounts(0:3),zrecvcounts(0:3)
    integer(4):: zdispls(0:3)
    integer(4):: mpierror
    include "mpif.h"
    phi = 0.

    call MPI_INIT(mpierror)
    call MPI_COMM_SIZE(MPI_COMM_WORLD,nzprocs,mpierror)
    call MPI_COMM_RANK(MPI_COMM_WORLD,izproc,mpierror)

    zsendcounts=0
    zrecvcounts=0
    zdispls=0
    zdispls=0
    zsendtypes=MPI_DOUBLE_PRECISION
    zrecvtypes=MPI_DOUBLE_PRECISION

    do ii=1,1000000000
      if (mod(ii,1000000_4) == 0) print*,"loop ",ii,izproc

      call MPI_ALLTOALLW(phi,zsendcounts,zdispls,zsendtypes,
   &                     phi2,zrecvcounts,zdispls,zrecvtypes,
   &                     MPI_COMM_WORLD,mpierror)

    enddo
    return
    end

_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users

I've been enhancing my code and have started using the nice routine alltoallw. The code works fine except that there seems to be a memory leak in alltoallw. I've eliminated all other possible causes and have reduced the code down to a bare minimum. I've included fortran source code which produces the problem. This code just keeps calling alltoallw, but with all of the send and receive counts set to zero, so it shouldn't be doing anything. And yet I can watch the memory continue to grow. As a sanity check, I change the code to call alltoallv instead, and there is no memory leak. If it helps, I am using OpenMPI on an AMD system running Chaos linux. I tried the latest nightly build of version 1.3 from Aug 5. I run four processors on one quad core node so it should be using shared memory communication.

smime.p7s
Description: S/MIME cryptographic signature