[petsc-users] Best way to scatter a Seq vector ?

Fri Jan 6 14:24:00 CST 2017

Great help Barry, i totally had overlooked that option (it is explicit in
the vecscatterbegin call help page but not in vecscattercreatetozero, as i
read later)

So i used that and it works partially, it scatters te values assigned in
root but not the rest, if i call vecscatterbegin from outside root it
hangs, the code currently look as this:

  call VecScatterCreateToZero(bp2,ctr,bp0,ierr); CHKERRQ(ierr)

  call PetscObjectSetName(bp0, 'bp0:',ierr)

if(rankl==0)then

     call VecSetValues(bp0,nbdp,ind,Rhs,INSERT_VALUES,ierr)

     call VecView(bp0,PETSC_VIEWER_STDOUT_WORLD,ierr)

     call VecScatterBegin(ctr,bp0,bp2,INSERT_VALUES,SCATTER_REVERSE,ierr)

     call VecScatterEnd(ctr,bp0,bp2,INSERT_VALUES,SCATTER_REVERSE,ierr)

     print*,"done! "

     CHKERRQ(ierr)

endif

   ! call VecScatterBegin(ctr,bp0,bp2,INSERT_VALUES,SCATTER_REVERSE,ierr)

   !  call VecScatterEnd(ctr,bp0,bp2,INSERT_VALUES,SCATTER_REVERSE,ierr)

  call VecView(bp2,PETSC_VIEWER_STDOUT_WORLD,ierr)

  call PetscBarrier(PETSC_NULL_OBJECT,ierr)

  call exit()

And the output is: (with bp the right answer)

Vec Object:bp: 2 MPI processes

  type: mpi

Process [0]

1.

2.

Process [1]

4.

3.

Vec Object:bp2: 2 MPI processes  *(before scatter)*

  type: mpi

Process [0]

0.

0.

Process [1]

0.

0.

Vec Object:bp0: 1 MPI processes

  type: seq

1.

2.

4.

3.

 done!

Vec Object:bp2: 2 MPI processes  *(after scatter)*

  type: mpi

Process [0]

1.

2.

*Process [1]*

*0.*

*0.*

Thanks inmensely for your help,

Manuel

On Thu, Jan 5, 2017 at 4:39 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:

>
> > On Jan 5, 2017, at 6:21 PM, Manuel Valera <mvalera at mail.sdsu.edu> wrote:
> >
> > Hello Devs is me again,
> >
> > I'm trying to distribute a vector to all called processes, the vector
> would be originally in root as a sequential vector and i would like to
> scatter it, what would the best call to do this ?
> >
> > I already know how to gather a distributed vector to root with
> VecScatterCreateToZero, this would be the inverse operation,
>
>    Use the same VecScatter object but with SCATTER_REVERSE, not you need
> to reverse the two vector arguments as well.
>
>
> > i'm currently trying with VecScatterCreate() and as of now im doing the
> following:
> >
> >
> > if(rank==0)then
> >
> >
> >      call VecCreate(PETSC_COMM_SELF,bp0,ierr); CHKERRQ(ierr) !if i use
> WORLD
> >                                                              !freezes in
> SetSizes
> >      call VecSetSizes(bp0,PETSC_DECIDE,nbdp,ierr); CHKERRQ(ierr)
> >      call VecSetType(bp0,VECSEQ,ierr)
> >      call VecSetFromOptions(bp0,ierr); CHKERRQ(ierr)
> >
> >
> >      call VecSetValues(bp0,nbdp,ind,Rhs,INSERT_VALUES,ierr)
> >
> >      !call VecSet(bp0,5.0D0,ierr); CHKERRQ(ierr)
> >
> >
> >      call VecView(bp0,PETSC_VIEWER_STDOUT_WORLD,ierr)
> >
> >      call VecAssemblyBegin(bp0,ierr) ; call VecAssemblyEnd(bp0,ierr) !rhs
> >
> >      do i=0,nbdp-1,1
> >         ind(i+1) = i
> >      enddo
> >
> >      call ISCreateGeneral(PETSC_COMM_SELF,nbdp,ind,PETSC_COPY_
> VALUES,locis,ierr)
> >
> >     !call VecScatterCreate(bp0,PETSC_NULL_OBJECT,bp2,is,ctr,ierr) !if i
> use SELF
> >
>  !freezes here.
> >
> >      call VecScatterCreate(bp0,locis,bp2,PETSC_NULL_OBJECT,ctr,ierr)
> >
> > endif
> >
> > bp2 being the receptor MPI vector to scatter to
> >
> > But it freezes in VecScatterCreate when trying to use more than one
> processor, what would be a better approach ?
> >
> >
> > Thanks once again,
> >
> > Manuel
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > On Wed, Jan 4, 2017 at 3:30 PM, Manuel Valera <mvalera at mail.sdsu.edu>
> wrote:
> > Thanks i had no idea how to debug and read those logs, that solved this
> issue at least (i was sending a message from root to everyone else, but
> trying to catch from everyone else including root)
> >
> > Until next time, many thanks,
> >
> > Manuel
> >
> > On Wed, Jan 4, 2017 at 3:23 PM, Matthew Knepley <knepley at gmail.com>
> wrote:
> > On Wed, Jan 4, 2017 at 5:21 PM, Manuel Valera <mvalera at mail.sdsu.edu>
> wrote:
> > I did a PetscBarrier just before calling the vicariate routine and im
> pretty sure im calling it from every processor, code looks like this:
> >
> > From the gdb trace.
> >
> >   Proc 0: Is in some MPI routine you call yourself, line 113
> >
> >   Proc 1: Is in VecCreate(), line 130
> >
> > You need to fix your communication code.
> >
> >    Matt
> >
> > call PetscBarrier(PETSC_NULL_OBJECT,ierr)
> >
> > print*,'entering POInit from',rank
> > !call exit()
> >
> > call PetscObjsInit()
> >
> >
> > And output gives:
> >
> >  entering POInit from           0
> >  entering POInit from           1
> >  entering POInit from           2
> >  entering POInit from           3
> >
> >
> > Still hangs in the same way,
> >
> > Thanks,
> >
> > Manuel
> >
> >
> >
> > On Wed, Jan 4, 2017 at 2:55 PM, Manuel Valera <mvalera at mail.sdsu.edu>
> wrote:
> > Thanks for the answers !
> >
> > heres the screenshot of what i got from bt in gdb (great hint in how to
> debug in petsc, didn't know that)
> >
> > I don't really know what to look at here,
> >
> > Thanks,
> >
> > Manuel
> >
> > On Wed, Jan 4, 2017 at 2:39 PM, Dave May <dave.mayhem23 at gmail.com>
> wrote:
> > Are you certain ALL ranks in PETSC_COMM_WORLD call these function(s).
> These functions cannot be inside if statements like
> > if (rank == 0){
> >   VecCreateMPI(...)
> > }
> >
> >
> > On Wed, 4 Jan 2017 at 23:34, Manuel Valera <mvalera at mail.sdsu.edu>
> wrote:
> > Thanks Dave for the quick answer, appreciate it,
> >
> > I just tried that and it didn't make a difference, any other suggestions
> ?
> >
> > Thanks,
> > Manuel
> >
> > On Wed, Jan 4, 2017 at 2:29 PM, Dave May <dave.mayhem23 at gmail.com>
> wrote:
> > You need to swap the order of your function calls.
> > Call VecSetSizes() before VecSetType()
> >
> > Thanks,
> >   Dave
> >
> >
> > On Wed, 4 Jan 2017 at 23:21, Manuel Valera <mvalera at mail.sdsu.edu>
> wrote:
> > Hello all, happy new year,
> >
> > I'm working on parallelizing my code, it worked and provided some
> results when i just called more than one processor, but created artifacts
> because i didn't need one image of the whole program in each processor,
> conflicting with each other.
> >
> > Since the pressure solver is the main part i need in parallel im chosing
> mpi to run everything in root processor until its time to solve for
> pressure, at this point im trying to create a distributed vector using
> either
> >
> >      call VecCreateMPI(PETSC_COMM_WORLD,PETSC_DECIDE,nbdp,xp,ierr)
> > or
> >      call VecCreate(PETSC_COMM_WORLD,xp,ierr); CHKERRQ(ierr)
> >      call VecSetType(xp,VECMPI,ierr)
> >      call VecSetSizes(xp,PETSC_DECIDE,nbdp,ierr); CHKERRQ(ierr)
> >
> >
> > In both cases program hangs at this point, something it never happened
> on the naive way i described before. I've made sure the global size, nbdp,
> is the same in every processor. What can be wrong?
> >
> > Thanks for your kind help,
> >
> > Manuel.
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
> > --
> > What most experimenters take for granted before they begin their
> experiments is infinitely more interesting than any results to which their
> experiments lead.
> > -- Norbert Wiener
> >
> >
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20170106/f209a7fc/attachment-0001.html>