Newsgroups: comp.parallel.mpi
From: gropp@godzilla.mcs.anl.gov (William Gropp)
Subject: Re: MPI_gatherv
Organization: MCS, Argonne National Laboratory
Date: Fri, 15 May 1998 17:44:4 GMT
Message-ID: <8952542441544@godzilla.mcs.anl.gov>

In article <355B0EB4.D72967C8@dl.ac.uk>, Ian Bush  <I.J.Bush@dl.ac.uk> wrote:
>Hi All,
>
>    I'm hoping that some standards lawyer might be able to confirm
>my understanding of something ...
>
>    The basic question is can you have sendcount=0 in MPI_gatherv,
>and if you can what corresponding values you should have for displs ?
>As I read it IF sendcount=0 is allowed, and it is not explicitly
>stated to be prohibited, then any old value for the relevant displs
>will do, e.g. 0, because of the sentence
>
>"The specification of counts, types and displacements should not cause 
>any location on the root to be written more than once"
>
>as if sendcount=0 for that processor no writing will occur.
>
>    I'm asking because for various perverse reasons I want to gather
>from a subset of the communicators processors. Yes, maybe I am being
>a bit lazy, MPI_comm_split to make the communicator with the interesting
>processors in it, gatherv then free the new communicator might be
>better, 
>but a new communicator for just one operation seems a little bit ugly to 
>me, hence the question,
>
>Ian

Yes, you should be able to do this.  In fact, in the MPI-2 standard, this
sort of approach is used as an illustation of how to achieve the "half duplex"
version of the intercommunicator collective operations (e.g., advice to
users for MPI_Alltoall, section 7.3.2, in MPI-2).

Bill