X-Mailer: Windows Eudora Pro Version 3.0.1 (32)
Date: Sun, 02 Aug 1998 15:52:17 -0500
To: Raja Daoud <raja@XXXXXXXXXXXXXXX>
From: William Gropp <gropp@XXXXXXXXXXX>
Subject: Re: MPI_ALLOC_MEM & no mem
Cc: treumann@XXXXXXXXXX,mpi-core@XXXXXXXXXXX,In-Reply-To: <199808012224.PAA07169@XXXXXXXXXXXXXX>
References: <5010400025483016000002L062*@XXX>
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Sender: owner-mpi-core@XXXXXXXXXXX
Precedence: bulk
X-UIDL: 9a57e0ea19b13cc79874a3c1b17ce8a3

One other thing to keep in mind: the reason ERRORS_ARE_FATAL is the default
is that users often fail to check error returns.  How many codes fail to
check the return from malloc?

The issue of whether the error is continuable has always been taken by the
MPI Forum as a quality-of-implementation issue.  A good implementation will
strive to make all errors continuable, modulo concerns on performance.  I'd
rather not get into the issue of how this impacts portability, other than
to remind everyone that any Fortran program that expects character*2 or an
open of unit=2 to work is also beyond what the standard specifies.  In C,
malloc could *always* return null.  Perhaps MPI is simply a little more up
front with the limitations.

Bill