From: Raja Daoud <raja@XXXXXXXXXXXXXXX>
Subject: Re: MPI_ALLOC_MEM & no mem
To: treumann@XXXXXXXXXX
Date: Sat, 01 Aug 1998 17:33:27 CDT
Cc: mpi-core@XXXXXXXXXXX,mpi-comments@XXXXXXXXXXXXX
In-Reply-To: <5010400025483016000002L062*@XXX>
X-Mailer: Elm [revision: 111.1]
Sender: owner-mpi-core@XXXXXXXXXXX
Precedence: bulk
X-UIDL: 4a72b40cc812385fee1614e9cb3ba023


Some approaches:

1- No change to MPI, implementors provide an option to set the amount of
   "special memory" needed.

2- Add a try_alloc routine.

3- Add a routine that allocates up to the given size and returns the
   actual size (avoid loop of try_alloc; lose flexibility in deciding
   the desired sizes).

4- Reconsider a tiny set of continuable errors.


I prefer #4 and believe it can be done in a manner to address the
concerns of the users and implementors who voted it down:

a- The default MPI_ERRORS_ARE_FATAL stays: users who are concerned
   about application error checking complexity are happy.

b- The small set of continuable errors is decided by consensus: only
   relevant cases are handled, and only if everybody agrees they can
   be trivially implemented (an if() very early inside the routine).

c- The small set is then mandated: portability concerns are addressed.

d- We have sufficient time to debate and evaluate till MPI 2.1.


With this MPI gains some flexibility for applications beyond the core of
scientific computing.  If (b) leads to the empty set, we're back to #1
with no additional harm done.  Comments?

--Raja

-=-
Raja Daoud				Hewlett-Packard Co.
raja@XXXXXXXXXX				http://www.hp.com/go/mpi