From: Karl Feind <kaf@XXXXXXX>
Subject: Re: Another clarification
To: treumann@XXXXXXXXXX
Date: Tue, 16 May 2000 13:09:42 -0500 (CDT)
Cc: Rabenseifner@XXXXXXXXXXXXXXXXXXXX,In-Reply-To: <OF779CFB09.9487AD44-ON852568DB.00619669@XXXXXXXXXXX>, 2000 01:51:16 PM
X-Mailer: ELM [version 2.5 PL2]
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit

Hi,

I think there are three relevant cases when executing MPI_Get_count and
computing (bytes_transferred/datatype_length):

	Case 1) bytes_transferred > 0 and datatype_length > 0
		This is the typical easy case.

	Case 2) bytes_transferred == datatype_length == 0
		This is Rolf's case and the proposal is to compute 0/0 = 0
	        as a special case.  This seems OK.

	Case 3) bytes_transferred > 0 and datatype_length == 0
		This seems erroneous and should cause MPI_UNDEFINED to 
		be returned.

What I've just described is how SGI MPI handles this.

Karl Feind

-----------
Karl Feind, Message-Passing Engineering Team           
Computer Systems Business Unit,  SGI
Phone: 651-683-5673  Fax: 651/683-5276  E-Mail: kaf@XXXXXXX          
		
	
treumann@XXXXXXXXXX writes:
> 
> 
> It  works for me as a user convenience in a case like this.   I can go with
> Zero. -  Thanks Rolf for the rationale.
> 
> Dick Treumann                IBM RS/6000 SP Development
> Tele (914) 433-7846         Fax (914) 433-8363
> 
> 
> Rabenseifner@XXXXXXXXXXXXXXXXXXXX (Rolf Rabenseifner) on 05/10/2000
> 10:47:06 AM
> 
> To:   Richard Treumann/Poughkeepsie/IBM@IBMUS
> cc:   William Gropp <gropp@XXXXXXXXXXX>, mpi-core@mcs.anl.gov> Subject:  Re: Another clarification
> 
> 
> 
> For the following user situation, I would prefer "0":
> - datatype is produced with  MPI...DARRAY and the definition of
>   DARRAY results in an empty block on some MPI process.
> - Therefore empty messages are transferred
> - The user use MPI_GET_COUNT to analyze the status.
> - He wants to make SPMD programming, i.e. without any
>   IF-statement, if not really necessary.
> 
> The clarification may also mention, that MPI_GET_COUNT will return "0"
> also if the sending count was positive.
> 
> Rolf
> 
> 
> Dick Treumann wrote:
> > Bill -
> >
> > You raise the issue of a zero byte datatype being used to query
> > MPI_Get_count or MPI_Get_elements on a status.  Both calls are allowed to
> > return MPI_UNDEFINED yet you suggest the answer zero be mandated.   We
> > detect the divide by 0 attempt and return MPI_UNDEFINED.  This seems like
> > the right result to me.
> >
> > Can you explain why you prefer zero?  I am interested if it is because
> zero
> > is better for some real user situation or because you consider zero  more
> > consistent with the MPI standard (or both).
> 
> 
> Bill Gropp wrote:
> >> The following question came up:  Suppose a datatype is created that has
> >> size 0; for example, a contiguous type with count 0 (the count is
> required
> >> to be non-negative, not positive).  A message is sent with this datatype
> >> and received, creating a status.  What value does MPI_Get_count or
> >> MPI_Get_elements return?  It could return 0, the count value specified
> by
> >> the user (just kidding!), MPI_UNDEFINED, or something else.
> >> I propose the clarification that in this case, MPI_Get_count and
> >> MPI_Get_elements return 0.
> 
> 
> Dr. Rolf Rabenseifner                      High Performance Computing
> Parallel Computing                         Center Stuttgart    (HLRS)
> Rechenzentrum Universitaet Stuttgart (RUS) Phone:    ++49 711 6855530
> Allmandring 30                             FAX:      ++49 711 6787626
> D-70550 Stuttgart                   rabenseifner@XXXXXXXXXXXXXXXXXXXX
> Germany                        http://www.hlrs.de/people/rabenseifner
> 
> 
>