Tony, Jim is making a correct observation. Would you please add his point to
the text?
> Marc,
>
> you give a set of arguments why even the "hard case" in place collective
> operations may be worth supporting. For completeness you might also
> give at least one argument against :-
>
> If the MPI implementation cannot re-order the sends and receives to
> ensure that data to be sent is sent before the same buffer locations
> need to be used for received data (and this may not always be
> possible, or may have unpleasant sequentialising effects on the whole
> collective algorithm), then the MPI library will have to allocate
> buffer space. Unless these operations are persistent, this buffer
> allocation/release will have to occur on every call. In effect this
> moves store allocation which could be cheap (since the user may well
> know the required buffer size at compile time, while the library
> cannot) and which would have occurred outside the loop into the
> loop. Therefore the use of in-place functions may actually turn out to
> be a pessimisation over the existing code using separate in and out
> buffers.
>
> -- Jim
>
> James Cownie
> BBN UK Ltd
> Phone : +44 117 9071438
> E-Mail: jcownie@bbn.com
>
-------------------
Marc Snir
IBM T.J. Watson Research Center
P.O. Box 218, Yorktown Heights, NY 10598
email: snir@watson.ibm.com
phone: 914-945-3204
fax: 914-945-4425
URL: http://www.research.ibm.com/xw-people-snir