[mpich-discuss] Problems with multiple calls to MPI_Reduce?

[Rajeev Thakur]

The collective operations don't use tags, so that should not be a problem.

I don't quite understand how you are using allreduce with maxloc to fit the
output together. 
 
You could use MPI-IO instead to write directly from each process to a common
file. Or a slower approach would be for each process to send the data to
rank 0 and have it write the file.

Rajeev 

> -----Original Message-----
> From: owner-mpich-discuss at mcs.anl.gov 
> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of Elliot Parkin
> Sent: Tuesday, July 01, 2008 4:32 AM
> To: mpich-discuss at mcs.anl.gov
> Subject: RE: [mpich-discuss] Problems with multiple calls to 
> MPI_Reduce?
> 
> Thanks for the reply guys, very much appreciated.
> 
> Just to set the scene of how and why I'm using MPI_Allreduce:
> 
> I'm using MPI to parallelize a hydrodynamics code. The hydro 
> grid is split into sections and to output a data file 
> containing the whole grid requires the separate parts to be 
> pieced together. The separate parts of the grid pass their 
> part to a dummy array and then MPI_Allreduce fits this 
> together and broadcasts the parts back to all processors. 
> Here's a chunk of the code:
> 
> c Multiple calls to glue for various parameters
> 
>           call glue(lzro,zro)    ! density
>           call glue(lzpr,zpr)    ! pressure
>           call glue(lzco,zco)    ! colour
>           call glue(lzux,zux)    ! x-velocity
>           call glue(lzuy,zuy)    ! y-velocity
>           call glue(lzuz,zuz)    ! z-velocity
> 
> c The glue subroutine looks like this:
> 
>        all = (imax*jmax*kmax)
> 
>        do k=1,kmax
>           do j=1,jmax
>              do i=1,imax
>                 dummy(i,j,k) = -1.0e50
>              end do
>           end do
>        end do
> 
>        do k=1,mpikmax
>           do j=1,mpijmax
>              do i=1,mpiimax
>                 a = j + rank*mpijmax
>                 dummy(i,a,k) = input(i,j,k)
>              end do
>           end do
>        end do
> 
>        call MPI_Allreduce (dummy,output,all,MPI_2DOUBLE_PRECISION,
>       :     MPI_MAXLOC,comm,ierr)
> 
> 
> Michael Ahlmann suggested that the problem could be due to 
> using non-unique tags for the calls. At the minute the calls 
> all use the same communicator, but I have to apologise about 
> my lack of knowledge about MPI because I don't really know if 
> that would be a problem or not.
> 
> Thanks again,
> 
> Elliot Parkin
> 
> 
> Quoting Rajeev Thakur <thakur at mcs.anl.gov>:
> 
> > Can you send us a small code fragment?
> >
> > Rajeev
> >
> >> -----Original Message-----
> >> From: owner-mpich-discuss at mcs.anl.gov 
> >> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of Elliot Parkin
> >> Sent: Friday, June 27, 2008 6:28 AM
> >> To: mpich-discuss at mcs.anl.gov
> >> Subject: [mpich-discuss] Problems with multiple calls to 
> MPI_Reduce?
> >>
> >> Hello everyone,
> >>
> >> I've ran into some difficulties using MPI_Reduce. It seems 
> that after 
> >> it has been called some of the parameters in the code get set to 
> >> bizarre values. Also, if I call MPI_Reduce consecutively 
> (to fill an 
> >> array with the maximum values across the processes) it 
> doesn't appear 
> >> to work the third time and sets all values of the array to 
> zero. Has 
> >> anybody else had this problem, and if so is there any way to 
> >> explain/fix this?
> >>
> >> Cheers,
> >>
> >> Elliot Parkin.
> >>
> >> ----------------------------------------------------------------
> >> This message was sent using IMP, the Internet Messaging Program.
> >>
> >>
> >
> >
> 
> 
>