[mpich-discuss] Query regarding structure of dynamic array derived datatype

[Jed Brown]

On Thu, Mar 21, 2013 at 8:29 PM, Pavan Balaji <balaji at mcs.anl.gov> wrote:

> You'll still need to create a datatype for MPI_Pack.  This is the
> structure you mentioned:
>
> struct A
> {
>     int size;
>     int *lptr;
> };
>
> I'm assuming you want to send the size and the data associated with lptr
> in a single message.  In this case, the only approaches I can think of are:
>
> 1. Create a datatype every time you want to send this information and
> free it once you are done.
>
> 2. Send two messages, one with the size parameter and the second with
> the data pointed to by lptr.
>
> Someone else might have a better way to do this.
>

If the receiver does not know the size in advance (or at least an upper
bound), then they would have to allocate before receiving anyway. If you
have to do that, you may as well either send two messages or use
MPI_Probe/MPI_Iprobe and allocate once you know the message size.

If the receiver knows an upper bound in advance, just MPI_Recv that upper
bound and use MPI_Get_count to find out how much you received.

If you have many of these, I would send all the sizes in one message and
all the data in another message, unless the receiver knows the sizes in
advance.

Once you know all the sizes, you could make an MPI_Type_create_hindexed and
send with respect to MPI_BOTTOM to send all those separate arrays without
explicit packing. I would only bother with that in a memory-constrained
environment or after exhausting the simpler options.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mpich.org/pipermail/discuss/attachments/20130321/b785ea5f/attachment.html>