Newsgroups: comp.parallel.mpi
From: Christophe Koudella <ckoudell@ens-lyon.fr>
Subject: topologies_MPI
Organization: Ens Lyon
Date: Wed, 19 Feb 1997 15:16:00 +0100
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <330B0B20.41C6@ens-lyon.fr>

Hello, 

I would like to know if anybody of you has experience with 
the function MPI_CART_CREATE ?

This function although rather useful , numbers orders the 
processors in the cartesian topology in 'C fashion', that is 
it orders them such that the last dimension of the communicator 
counts the fastest. 

However in the code I am writing I am also using BLACS routines and 
there the call to gridinit also creates a cartesian topology , but 
this time it numbers the processors in fortran fashion, that is the 
first coordinate counts the fastest. 

I do however want compatibility between the two topologies . If I 
refer to the rank of some processor in a routine say, then of course 
the coordinate information given by BLACS disagrees with that given 
by MPI. Bad really.. 

So a way out is to reorder the processors as far as MPI is concerned, 
before calling the MPI_CART_CREATE routine. The routine is called 
with the new processor order and the numbering is of course still 
in C fashion, but the old order of the processors now may be brougth 
to agreement with the BLACS topology. 

Anyway I would be quite interested in your experience, because I am 
worried that this method makes things slow down. 

bye,bye 

Christophe