Jeff
Thanks for all this help. I really appreciate it. For various
reasons now different nodes don't just have 1 continguous piece
of array to be allgather'd. All nodes have multiple contiguous
pieces spread out over the array.
Is it possible to define an MPI datatype, like you suggested,
to handle this general case or do you recommend another strategy??
Chris
On Mon, Sep 22, 2003 at 07:21:37AM -0400, Jeff Squyres wrote:
> You can actually do it either day (using the language's native ordering
> one way or MPI datatypes the other day).
>
> C's native ordering is row-major; Fortran's native ordering is
> column-major.
>
> It looks like you want to gather rows of a Fortran array, so you'll need
> to make an MPI datatype that contains a row (i.e., a(1,1-3)), and then
> you'll allgather 1 of those (since every MPI process will be contributing
> 1 row to the collective).
>
>
>
> On Sun, 21 Sep 2003 seberino_at_[hidden] wrote:
>
> > Jeff
> >
> > I hope you are right that you can do mpi_allgatherv's
> > just as easily with 2d arrays as with 1d arrays.
> >
> > I'm confused however why you don't see a gotcha regarding
> > order of indices...
> >
> > "3 elements from node0" means different things for different
> > orderings of indices don't you agree?....
> >
> > It may mean {a(1,1),a(2,1),a(3,1)} or {a(1,1),a(1,2),a(1,3)}.
> > You'll "gather" different guys from same nodes!!!!
> >
> > Do you see what I mean?
> >
> > chris
> >
> > On Mon, Aug 04, 2003 at 05:20:46PM -0400, Jeff Squyres wrote:
> > > On Sat, 2 Aug 2003 seberino_at_[hidden] wrote:
> > >
> > > > Is an allgather possible to give every node a complete copy of this 2D
> > > > array that is spread out over all the nodes???
> > > >
> > > > Node Number What part of array "a" it has
> > > > =================================================
> > > >
> > > > 0 a(1,1), a(1,2), a(1,3) (i.e. a(1,1-3))
> > > > 1 a(2,1), a(2,2), a(2,3) (i.e. a(2,1-3))
> > > > 2 a(3,1), a(3,2), a(3,3) (i.e. a(3,1-3))
> > > > 3 a(4,1), a(4,2), a(4,3) (i.e. a(4,1-3))
> > > >
> > > > 4 a(1,4), a(1,5), a(1,6) (i.e. a(1,4-6))
> > > > 5 a(2,4), a(2,5), a(2,6) (i.e. a(2,4-6))
> > > > 6 a(3,4), a(3,5), a(3,6) (i.e. a(3,4-6))
> > > > 7 a(4,4), a(4,5), a(4,6) (i.e. a(4,4-6))
> > >
> > > Yes, it sounds like an Allgather is what you want to do. Check out Figure
> > > 4.1 in the MPI-1 spec (p92) for a handy diagram of what each of the data
> > > movement collectives do in MPI.
> > >
> > > > I've heard about column-major order and row-major order but always
> > > > forget which one is C and which one Fortran90 uses.
>
> --
> {+} Jeff Squyres
> {+} jsquyres_at_[hidden]
> {+} http://www.lam-mpi.org/
> _______________________________________________
> This list is archived at http://www.lam-mpi.org/MailArchives/lam/
--
_______________________________________
Christian Seberino, Ph.D.
SPAWAR Systems Center San Diego
Code 2872
49258 Mills Street, Room 158
San Diego, CA 92152-5385
U.S.A.
Phone: (619) 553-9973
Fax : (619) 553-6521
Email: seberino_at_[hidden]
_______________________________________
|
