Newsgroups: comp.parallel.mpi,comp.unix.cray
From: thimm@physik.fu-berlin.de (Axel Thimm)
Subject: MPI Non blocking Sends, send buffer not readable or unclear standard?
Organization: FU Berlin, Fachbereich Physik, Institut fuer Theoretische Physik
Date: Sat, 25 May 1996 12:03:06 GMT
Message-ID: <4o6stc$e1q@fu-berlin.de>

Hello,

When using non blocking communication, the standard (MPI Report 1.1)
says that "The sender may not access any part of the send buffer after a
nonblocking send operation is called, until the send completes." As I
understand it, the "send buffer" is either the memory location for
standard/ready/sync modes or the internal MPI buffer when using buffered
mode communication.
Is there any reason that in the former case the send buffer should not
be read again?
I would like to use ready mode nonbl. communications, where the data
should be send to multiple processes and also used in subsequent
computations (read only). The send completion should be forced only when
the owner process needs to update the data. I can of course use buffered
non blocking sends to avoid this problem, but this will cause enormous
overheads.
If this is implementation specific, I would like to know how the EPCC
MPI impl. for the T3D behaves.

Thank you in advance for any responses (please use (also) email if
possible).
Regards,
	Axel Thimm.
--
Axel Thimm <thimm@physik.fu-berlin.de>
Freie Univ. Berlin, Fachbereich Physik, Inst. fuer Theor. Physik
Arnimallee 14, D-14195 Berlin, Raum 0.4.08, Tel: 838-4774


