Newsgroups: comp.parallel.mpi
From: EDDY <eddy@lseet.univ-tln.fr>
Subject: Re: Help needed,please II
Organization: Universite de Toulon et du Var
Date: Wed, 07 Jan 1998 14:34:37 +0100
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <34B3846D.3981541D@lseet.univ-tln.fr>

> >Sender:                Receiver:
> 
> >1  7   13             1    2    3    4   5   6
> >2  8   14             7    8    9   10  11  17
> >3  9   15             13   14   15  16  17  18
> >4  10  16             .    .    .   .   .   .
> >5  11  17             .    .    .   .   .   .
> >6  12  18
> >.  .   .
> >.  .   .
> 
> Receiver : use the data structure suggested earlier, with number of
> blocks 6 (in this case), 1 element per block, stride = leading
> dimension of array at receiver side. Lower bound = 0, upper bound = 1.
> Receive (in this case) 3 rows.
> 
> Sender : create a data type for 6 contiguous data elements, set lower
> bound to 0, upper bound to leading dimension of the array at sender
> side. Send 3 columns.
> 
> As before, this scheme uses only 1 datatype, as long as your
> sendercolumnlength/receiverrowlength remains the same ; the
> sendernumberofcolumns/receivernumberofrows can be selected by changing
> the count in the send/receive.

Yes, it will be work. But if sender must send *different* number of its
rows too, it seems again the old problem will be arised :( .

I think now, that to decide this problem some missing in MPI 
capabilities  are neccesary.