[petsc-users] petsc4py and MPI.COMM_SELF.Spawn

[Matthew Knepley]

On Tue, Feb 28, 2017 at 6:17 PM, Rodrigo Felicio <Rodrigo.Felicio at iongeo.com
> wrote:

> Dear All,
>
> I am new to both pestc (petsc4py) and MPI. I am prototyping  an
> application that acts on data belonging to thousands of different spatial
> locations, so I decided to use mpi4py to parallelize my computations over
> said  set of points. At each point, however, I have to solve  linear
> systems of the type A x = b, using least-squares. As matrices A are sparse
> and large, I would like use petsc LSQR solver to obtain x. Trouble is that
> I cannot get mpi4py to work with petsc4py when using MPI.Spawn(), so that I
> could  solve those linear systems using threads spawned from my main
> algorithm. I am not sure if this is a problem with my petsc/mpi
> installation, or if there is some inherent incompatibility between mpi4py
> and petsc4py on spawned threads.  Anyway, I would be really  thankful to
> anyone who could shed some light on this issue. For illustration,  the
> child and parent codes found on the pestc4py spawning demo folder fail for
> me if I just include the petsc4py initialization in either of them. In
> fact, just the introduction of the line " from petsc4py import PETSc" is
> enough to make the code to hang and issue, upon keyboard termination, error
> msgs of the type "spawned process group was unable to connect back to
> parent port" .
>

MPI is really an SPMD framework, so it does not make sense to me that you
separate the codes here. Normally
you would designate one rank (say 0) to be the master, and then the others
are the children. Also, startup is
collective, so everyone must call it at the same time.

  Matt


> Kind regards
> Rodrigo
>
> ------------------------
>  CHILD code:
> -------------------------
>
> # inclusion of petsc4py lines causes code to hang
> import petsc4py
> petsc4py.init(sys.argv)
> from petsc4py import PETSc
> from mpi4py import MPI
> from array import array
>
> master = MPI.Comm.Get_parent()
> nprocs = master.Get_size()
> myrank = master.Get_rank()
>
> n  = array('i', [0])
> master.Bcast([n, MPI.INT], root=0)
> n = n[0]
>
> h = 1.0 / n
> s = 0.0
> for i in range(myrank+1, n+1, nprocs):
>     x = h * (i - 0.5)
>     s += 4.0 / (1.0 + x**2)
> pi = s * h
>
> pi = array('d', [pi])
> master.Reduce(sendbuf=[pi, MPI.DOUBLE],
>               recvbuf=None,
>               op=MPI.SUM, root=0)
>
> master.Disconnect()
>
>
> ---------------------
> MASTER CODE:
> ----------------------
> from mpi4py import MPI
> from array import array
> from math import pi as PI
> from sys import argv
>
> cmd = 'cpi-worker-py.exe'
> if len(argv) > 1: cmd = argv[1]
> print("%s -> %s" % (argv[0], cmd))
>
> worker = MPI.COMM_SELF.Spawn(cmd, None, 5)
>
> n  = array('i', [100])
> worker.Bcast([n,MPI.INT], root=MPI.ROOT)
>
> pi = array('d', [0.0])
> worker.Reduce(sendbuf=None,
>               recvbuf=[pi, MPI.DOUBLE],
>               op=MPI.SUM, root=MPI.ROOT)
> pi = pi[0]
>
> worker.Disconnect()
>
> print('pi: %.16f, error: %.16f' % (pi, abs(PI-pi)))
>
>
>
>
>
> ________________________________
>
>
> This email and any files transmitted with it are confidential and are
> intended solely for the use of the individual or entity to whom they are
> addressed. If you are not the original recipient or the person responsible
> for delivering the email to the intended recipient, be advised that you
> have received this email in error, and that any use, dissemination,
> forwarding, printing, or copying of this email is strictly prohibited. If
> you received this email in error, please immediately notify the sender and
> delete the original.
>
>


-- 
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/20170301/46fc44d2/attachment.html>